uma_core.c revision 260305
1/*- 2 * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson <jeff@FreeBSD.org> 3 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org> 4 * Copyright (c) 2004-2006 Robert N. M. Watson 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice unmodified, this list of conditions, and the following 12 * disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 */ 28 29/* 30 * uma_core.c Implementation of the Universal Memory allocator 31 * 32 * This allocator is intended to replace the multitude of similar object caches 33 * in the standard FreeBSD kernel. The intent is to be flexible as well as 34 * effecient. A primary design goal is to return unused memory to the rest of 35 * the system. This will make the system as a whole more flexible due to the 36 * ability to move memory to subsystems which most need it instead of leaving 37 * pools of reserved memory unused. 38 * 39 * The basic ideas stem from similar slab/zone based allocators whose algorithms 40 * are well known. 41 * 42 */ 43 44/* 45 * TODO: 46 * - Improve memory usage for large allocations 47 * - Investigate cache size adjustments 48 */ 49 50#include <sys/cdefs.h> 51__FBSDID("$FreeBSD: stable/10/sys/vm/uma_core.c 260305 2014-01-04 23:42:24Z mav $"); 52 53/* I should really use ktr.. */ 54/* 55#define UMA_DEBUG 1 56#define UMA_DEBUG_ALLOC 1 57#define UMA_DEBUG_ALLOC_1 1 58*/ 59 60#include "opt_ddb.h" 61#include "opt_param.h" 62#include "opt_vm.h" 63 64#include <sys/param.h> 65#include <sys/systm.h> 66#include <sys/bitset.h> 67#include <sys/kernel.h> 68#include <sys/types.h> 69#include <sys/queue.h> 70#include <sys/malloc.h> 71#include <sys/ktr.h> 72#include <sys/lock.h> 73#include <sys/sysctl.h> 74#include <sys/mutex.h> 75#include <sys/proc.h> 76#include <sys/rwlock.h> 77#include <sys/sbuf.h> 78#include <sys/sched.h> 79#include <sys/smp.h> 80#include <sys/vmmeter.h> 81 82#include <vm/vm.h> 83#include <vm/vm_object.h> 84#include <vm/vm_page.h> 85#include <vm/vm_pageout.h> 86#include <vm/vm_param.h> 87#include <vm/vm_map.h> 88#include <vm/vm_kern.h> 89#include <vm/vm_extern.h> 90#include <vm/uma.h> 91#include <vm/uma_int.h> 92#include <vm/uma_dbg.h> 93 94#include <ddb/ddb.h> 95 96#ifdef DEBUG_MEMGUARD 97#include <vm/memguard.h> 98#endif 99 100/* 101 * This is the zone and keg from which all zones are spawned. The idea is that 102 * even the zone & keg heads are allocated from the allocator, so we use the 103 * bss section to bootstrap us. 104 */ 105static struct uma_keg masterkeg; 106static struct uma_zone masterzone_k; 107static struct uma_zone masterzone_z; 108static uma_zone_t kegs = &masterzone_k; 109static uma_zone_t zones = &masterzone_z; 110 111/* This is the zone from which all of uma_slab_t's are allocated. */ 112static uma_zone_t slabzone; 113static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */ 114 115/* 116 * The initial hash tables come out of this zone so they can be allocated 117 * prior to malloc coming up. 118 */ 119static uma_zone_t hashzone; 120 121/* The boot-time adjusted value for cache line alignment. */ 122int uma_align_cache = 64 - 1; 123 124static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); 125 126/* 127 * Are we allowed to allocate buckets? 128 */ 129static int bucketdisable = 1; 130 131/* Linked list of all kegs in the system */ 132static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); 133 134/* This mutex protects the keg list */ 135static struct mtx_padalign uma_mtx; 136 137/* Linked list of boot time pages */ 138static LIST_HEAD(,uma_slab) uma_boot_pages = 139 LIST_HEAD_INITIALIZER(uma_boot_pages); 140 141/* This mutex protects the boot time pages list */ 142static struct mtx_padalign uma_boot_pages_mtx; 143 144/* Is the VM done starting up? */ 145static int booted = 0; 146#define UMA_STARTUP 1 147#define UMA_STARTUP2 2 148 149/* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */ 150static const u_int uma_max_ipers = SLAB_SETSIZE; 151 152/* 153 * Only mbuf clusters use ref zones. Just provide enough references 154 * to support the one user. New code should not use the ref facility. 155 */ 156static const u_int uma_max_ipers_ref = PAGE_SIZE / MCLBYTES; 157 158/* 159 * This is the handle used to schedule events that need to happen 160 * outside of the allocation fast path. 161 */ 162static struct callout uma_callout; 163#define UMA_TIMEOUT 20 /* Seconds for callout interval. */ 164 165/* 166 * This structure is passed as the zone ctor arg so that I don't have to create 167 * a special allocation function just for zones. 168 */ 169struct uma_zctor_args { 170 const char *name; 171 size_t size; 172 uma_ctor ctor; 173 uma_dtor dtor; 174 uma_init uminit; 175 uma_fini fini; 176 uma_import import; 177 uma_release release; 178 void *arg; 179 uma_keg_t keg; 180 int align; 181 uint32_t flags; 182}; 183 184struct uma_kctor_args { 185 uma_zone_t zone; 186 size_t size; 187 uma_init uminit; 188 uma_fini fini; 189 int align; 190 uint32_t flags; 191}; 192 193struct uma_bucket_zone { 194 uma_zone_t ubz_zone; 195 char *ubz_name; 196 int ubz_entries; /* Number of items it can hold. */ 197 int ubz_maxsize; /* Maximum allocation size per-item. */ 198}; 199 200/* 201 * Compute the actual number of bucket entries to pack them in power 202 * of two sizes for more efficient space utilization. 203 */ 204#define BUCKET_SIZE(n) \ 205 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) 206 207#define BUCKET_MAX BUCKET_SIZE(128) 208 209struct uma_bucket_zone bucket_zones[] = { 210 { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, 211 { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, 212 { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, 213 { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, 214 { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, 215 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, 216 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, 217 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, 218 { NULL, NULL, 0} 219}; 220 221/* 222 * Flags and enumerations to be passed to internal functions. 223 */ 224enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; 225 226/* Prototypes.. */ 227 228static void *noobj_alloc(uma_zone_t, int, uint8_t *, int); 229static void *page_alloc(uma_zone_t, int, uint8_t *, int); 230static void *startup_alloc(uma_zone_t, int, uint8_t *, int); 231static void page_free(void *, int, uint8_t); 232static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); 233static void cache_drain(uma_zone_t); 234static void bucket_drain(uma_zone_t, uma_bucket_t); 235static void bucket_cache_drain(uma_zone_t zone); 236static int keg_ctor(void *, int, void *, int); 237static void keg_dtor(void *, int, void *); 238static int zone_ctor(void *, int, void *, int); 239static void zone_dtor(void *, int, void *); 240static int zero_init(void *, int, int); 241static void keg_small_init(uma_keg_t keg); 242static void keg_large_init(uma_keg_t keg); 243static void zone_foreach(void (*zfunc)(uma_zone_t)); 244static void zone_timeout(uma_zone_t zone); 245static int hash_alloc(struct uma_hash *); 246static int hash_expand(struct uma_hash *, struct uma_hash *); 247static void hash_free(struct uma_hash *hash); 248static void uma_timeout(void *); 249static void uma_startup3(void); 250static void *zone_alloc_item(uma_zone_t, void *, int); 251static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); 252static void bucket_enable(void); 253static void bucket_init(void); 254static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); 255static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); 256static void bucket_zone_drain(void); 257static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags); 258static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); 259static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); 260static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); 261static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); 262static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, 263 uma_fini fini, int align, uint32_t flags); 264static int zone_import(uma_zone_t zone, void **bucket, int max, int flags); 265static void zone_release(uma_zone_t zone, void **bucket, int cnt); 266 267void uma_print_zone(uma_zone_t); 268void uma_print_stats(void); 269static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); 270static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); 271 272SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 273 274SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 275 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); 276 277SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 278 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); 279 280static int zone_warnings = 1; 281TUNABLE_INT("vm.zone_warnings", &zone_warnings); 282SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RW, &zone_warnings, 0, 283 "Warn when UMA zones becomes full"); 284 285/* 286 * This routine checks to see whether or not it's safe to enable buckets. 287 */ 288static void 289bucket_enable(void) 290{ 291 bucketdisable = vm_page_count_min(); 292} 293 294/* 295 * Initialize bucket_zones, the array of zones of buckets of various sizes. 296 * 297 * For each zone, calculate the memory required for each bucket, consisting 298 * of the header and an array of pointers. 299 */ 300static void 301bucket_init(void) 302{ 303 struct uma_bucket_zone *ubz; 304 int size; 305 int i; 306 307 for (i = 0, ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { 308 size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 309 size += sizeof(void *) * ubz->ubz_entries; 310 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 311 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 312 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET); 313 } 314} 315 316/* 317 * Given a desired number of entries for a bucket, return the zone from which 318 * to allocate the bucket. 319 */ 320static struct uma_bucket_zone * 321bucket_zone_lookup(int entries) 322{ 323 struct uma_bucket_zone *ubz; 324 325 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 326 if (ubz->ubz_entries >= entries) 327 return (ubz); 328 ubz--; 329 return (ubz); 330} 331 332static int 333bucket_select(int size) 334{ 335 struct uma_bucket_zone *ubz; 336 337 ubz = &bucket_zones[0]; 338 if (size > ubz->ubz_maxsize) 339 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); 340 341 for (; ubz->ubz_entries != 0; ubz++) 342 if (ubz->ubz_maxsize < size) 343 break; 344 ubz--; 345 return (ubz->ubz_entries); 346} 347 348static uma_bucket_t 349bucket_alloc(uma_zone_t zone, void *udata, int flags) 350{ 351 struct uma_bucket_zone *ubz; 352 uma_bucket_t bucket; 353 354 /* 355 * This is to stop us from allocating per cpu buckets while we're 356 * running out of vm.boot_pages. Otherwise, we would exhaust the 357 * boot pages. This also prevents us from allocating buckets in 358 * low memory situations. 359 */ 360 if (bucketdisable) 361 return (NULL); 362 /* 363 * To limit bucket recursion we store the original zone flags 364 * in a cookie passed via zalloc_arg/zfree_arg. This allows the 365 * NOVM flag to persist even through deep recursions. We also 366 * store ZFLAG_BUCKET once we have recursed attempting to allocate 367 * a bucket for a bucket zone so we do not allow infinite bucket 368 * recursion. This cookie will even persist to frees of unused 369 * buckets via the allocation path or bucket allocations in the 370 * free path. 371 */ 372 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 373 udata = (void *)(uintptr_t)zone->uz_flags; 374 else { 375 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) 376 return (NULL); 377 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); 378 } 379 if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) 380 flags |= M_NOVM; 381 ubz = bucket_zone_lookup(zone->uz_count); 382 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); 383 if (bucket) { 384#ifdef INVARIANTS 385 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 386#endif 387 bucket->ub_cnt = 0; 388 bucket->ub_entries = ubz->ubz_entries; 389 } 390 391 return (bucket); 392} 393 394static void 395bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) 396{ 397 struct uma_bucket_zone *ubz; 398 399 KASSERT(bucket->ub_cnt == 0, 400 ("bucket_free: Freeing a non free bucket.")); 401 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) 402 udata = (void *)(uintptr_t)zone->uz_flags; 403 ubz = bucket_zone_lookup(bucket->ub_entries); 404 uma_zfree_arg(ubz->ubz_zone, bucket, udata); 405} 406 407static void 408bucket_zone_drain(void) 409{ 410 struct uma_bucket_zone *ubz; 411 412 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 413 zone_drain(ubz->ubz_zone); 414} 415 416static void 417zone_log_warning(uma_zone_t zone) 418{ 419 static const struct timeval warninterval = { 300, 0 }; 420 421 if (!zone_warnings || zone->uz_warning == NULL) 422 return; 423 424 if (ratecheck(&zone->uz_ratecheck, &warninterval)) 425 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); 426} 427 428static void 429zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) 430{ 431 uma_klink_t klink; 432 433 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) 434 kegfn(klink->kl_keg); 435} 436 437/* 438 * Routine called by timeout which is used to fire off some time interval 439 * based calculations. (stats, hash size, etc.) 440 * 441 * Arguments: 442 * arg Unused 443 * 444 * Returns: 445 * Nothing 446 */ 447static void 448uma_timeout(void *unused) 449{ 450 bucket_enable(); 451 zone_foreach(zone_timeout); 452 453 /* Reschedule this event */ 454 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 455} 456 457/* 458 * Routine to perform timeout driven calculations. This expands the 459 * hashes and does per cpu statistics aggregation. 460 * 461 * Returns nothing. 462 */ 463static void 464keg_timeout(uma_keg_t keg) 465{ 466 467 KEG_LOCK(keg); 468 /* 469 * Expand the keg hash table. 470 * 471 * This is done if the number of slabs is larger than the hash size. 472 * What I'm trying to do here is completely reduce collisions. This 473 * may be a little aggressive. Should I allow for two collisions max? 474 */ 475 if (keg->uk_flags & UMA_ZONE_HASH && 476 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { 477 struct uma_hash newhash; 478 struct uma_hash oldhash; 479 int ret; 480 481 /* 482 * This is so involved because allocating and freeing 483 * while the keg lock is held will lead to deadlock. 484 * I have to do everything in stages and check for 485 * races. 486 */ 487 newhash = keg->uk_hash; 488 KEG_UNLOCK(keg); 489 ret = hash_alloc(&newhash); 490 KEG_LOCK(keg); 491 if (ret) { 492 if (hash_expand(&keg->uk_hash, &newhash)) { 493 oldhash = keg->uk_hash; 494 keg->uk_hash = newhash; 495 } else 496 oldhash = newhash; 497 498 KEG_UNLOCK(keg); 499 hash_free(&oldhash); 500 return; 501 } 502 } 503 KEG_UNLOCK(keg); 504} 505 506static void 507zone_timeout(uma_zone_t zone) 508{ 509 510 zone_foreach_keg(zone, &keg_timeout); 511} 512 513/* 514 * Allocate and zero fill the next sized hash table from the appropriate 515 * backing store. 516 * 517 * Arguments: 518 * hash A new hash structure with the old hash size in uh_hashsize 519 * 520 * Returns: 521 * 1 on sucess and 0 on failure. 522 */ 523static int 524hash_alloc(struct uma_hash *hash) 525{ 526 int oldsize; 527 int alloc; 528 529 oldsize = hash->uh_hashsize; 530 531 /* We're just going to go to a power of two greater */ 532 if (oldsize) { 533 hash->uh_hashsize = oldsize * 2; 534 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 535 hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 536 M_UMAHASH, M_NOWAIT); 537 } else { 538 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 539 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, 540 M_WAITOK); 541 hash->uh_hashsize = UMA_HASH_SIZE_INIT; 542 } 543 if (hash->uh_slab_hash) { 544 bzero(hash->uh_slab_hash, alloc); 545 hash->uh_hashmask = hash->uh_hashsize - 1; 546 return (1); 547 } 548 549 return (0); 550} 551 552/* 553 * Expands the hash table for HASH zones. This is done from zone_timeout 554 * to reduce collisions. This must not be done in the regular allocation 555 * path, otherwise, we can recurse on the vm while allocating pages. 556 * 557 * Arguments: 558 * oldhash The hash you want to expand 559 * newhash The hash structure for the new table 560 * 561 * Returns: 562 * Nothing 563 * 564 * Discussion: 565 */ 566static int 567hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 568{ 569 uma_slab_t slab; 570 int hval; 571 int i; 572 573 if (!newhash->uh_slab_hash) 574 return (0); 575 576 if (oldhash->uh_hashsize >= newhash->uh_hashsize) 577 return (0); 578 579 /* 580 * I need to investigate hash algorithms for resizing without a 581 * full rehash. 582 */ 583 584 for (i = 0; i < oldhash->uh_hashsize; i++) 585 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 586 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 587 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 588 hval = UMA_HASH(newhash, slab->us_data); 589 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 590 slab, us_hlink); 591 } 592 593 return (1); 594} 595 596/* 597 * Free the hash bucket to the appropriate backing store. 598 * 599 * Arguments: 600 * slab_hash The hash bucket we're freeing 601 * hashsize The number of entries in that hash bucket 602 * 603 * Returns: 604 * Nothing 605 */ 606static void 607hash_free(struct uma_hash *hash) 608{ 609 if (hash->uh_slab_hash == NULL) 610 return; 611 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 612 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); 613 else 614 free(hash->uh_slab_hash, M_UMAHASH); 615} 616 617/* 618 * Frees all outstanding items in a bucket 619 * 620 * Arguments: 621 * zone The zone to free to, must be unlocked. 622 * bucket The free/alloc bucket with items, cpu queue must be locked. 623 * 624 * Returns: 625 * Nothing 626 */ 627 628static void 629bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 630{ 631 int i; 632 633 if (bucket == NULL) 634 return; 635 636 if (zone->uz_fini) 637 for (i = 0; i < bucket->ub_cnt; i++) 638 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); 639 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); 640 bucket->ub_cnt = 0; 641} 642 643/* 644 * Drains the per cpu caches for a zone. 645 * 646 * NOTE: This may only be called while the zone is being turn down, and not 647 * during normal operation. This is necessary in order that we do not have 648 * to migrate CPUs to drain the per-CPU caches. 649 * 650 * Arguments: 651 * zone The zone to drain, must be unlocked. 652 * 653 * Returns: 654 * Nothing 655 */ 656static void 657cache_drain(uma_zone_t zone) 658{ 659 uma_cache_t cache; 660 int cpu; 661 662 /* 663 * XXX: It is safe to not lock the per-CPU caches, because we're 664 * tearing down the zone anyway. I.e., there will be no further use 665 * of the caches at this point. 666 * 667 * XXX: It would good to be able to assert that the zone is being 668 * torn down to prevent improper use of cache_drain(). 669 * 670 * XXX: We lock the zone before passing into bucket_cache_drain() as 671 * it is used elsewhere. Should the tear-down path be made special 672 * there in some form? 673 */ 674 CPU_FOREACH(cpu) { 675 cache = &zone->uz_cpu[cpu]; 676 bucket_drain(zone, cache->uc_allocbucket); 677 bucket_drain(zone, cache->uc_freebucket); 678 if (cache->uc_allocbucket != NULL) 679 bucket_free(zone, cache->uc_allocbucket, NULL); 680 if (cache->uc_freebucket != NULL) 681 bucket_free(zone, cache->uc_freebucket, NULL); 682 cache->uc_allocbucket = cache->uc_freebucket = NULL; 683 } 684 ZONE_LOCK(zone); 685 bucket_cache_drain(zone); 686 ZONE_UNLOCK(zone); 687} 688 689static void 690cache_shrink(uma_zone_t zone) 691{ 692 693 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 694 return; 695 696 ZONE_LOCK(zone); 697 zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; 698 ZONE_UNLOCK(zone); 699} 700 701static void 702cache_drain_safe_cpu(uma_zone_t zone) 703{ 704 uma_cache_t cache; 705 uma_bucket_t b1, b2; 706 707 if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 708 return; 709 710 b1 = b2 = NULL; 711 ZONE_LOCK(zone); 712 critical_enter(); 713 cache = &zone->uz_cpu[curcpu]; 714 if (cache->uc_allocbucket) { 715 if (cache->uc_allocbucket->ub_cnt != 0) 716 LIST_INSERT_HEAD(&zone->uz_buckets, 717 cache->uc_allocbucket, ub_link); 718 else 719 b1 = cache->uc_allocbucket; 720 cache->uc_allocbucket = NULL; 721 } 722 if (cache->uc_freebucket) { 723 if (cache->uc_freebucket->ub_cnt != 0) 724 LIST_INSERT_HEAD(&zone->uz_buckets, 725 cache->uc_freebucket, ub_link); 726 else 727 b2 = cache->uc_freebucket; 728 cache->uc_freebucket = NULL; 729 } 730 critical_exit(); 731 ZONE_UNLOCK(zone); 732 if (b1) 733 bucket_free(zone, b1, NULL); 734 if (b2) 735 bucket_free(zone, b2, NULL); 736} 737 738/* 739 * Safely drain per-CPU caches of a zone(s) to alloc bucket. 740 * This is an expensive call because it needs to bind to all CPUs 741 * one by one and enter a critical section on each of them in order 742 * to safely access their cache buckets. 743 * Zone lock must not be held on call this function. 744 */ 745static void 746cache_drain_safe(uma_zone_t zone) 747{ 748 int cpu; 749 750 /* 751 * Polite bucket sizes shrinking was not enouth, shrink aggressively. 752 */ 753 if (zone) 754 cache_shrink(zone); 755 else 756 zone_foreach(cache_shrink); 757 758 CPU_FOREACH(cpu) { 759 thread_lock(curthread); 760 sched_bind(curthread, cpu); 761 thread_unlock(curthread); 762 763 if (zone) 764 cache_drain_safe_cpu(zone); 765 else 766 zone_foreach(cache_drain_safe_cpu); 767 } 768 thread_lock(curthread); 769 sched_unbind(curthread); 770 thread_unlock(curthread); 771} 772 773/* 774 * Drain the cached buckets from a zone. Expects a locked zone on entry. 775 */ 776static void 777bucket_cache_drain(uma_zone_t zone) 778{ 779 uma_bucket_t bucket; 780 781 /* 782 * Drain the bucket queues and free the buckets, we just keep two per 783 * cpu (alloc/free). 784 */ 785 while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 786 LIST_REMOVE(bucket, ub_link); 787 ZONE_UNLOCK(zone); 788 bucket_drain(zone, bucket); 789 bucket_free(zone, bucket, NULL); 790 ZONE_LOCK(zone); 791 } 792 793 /* 794 * Shrink further bucket sizes. Price of single zone lock collision 795 * is probably lower then price of global cache drain. 796 */ 797 if (zone->uz_count > zone->uz_count_min) 798 zone->uz_count--; 799} 800 801static void 802keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) 803{ 804 uint8_t *mem; 805 int i; 806 uint8_t flags; 807 808 mem = slab->us_data; 809 flags = slab->us_flags; 810 i = start; 811 if (keg->uk_fini != NULL) { 812 for (i--; i > -1; i--) 813 keg->uk_fini(slab->us_data + (keg->uk_rsize * i), 814 keg->uk_size); 815 } 816 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 817 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 818#ifdef UMA_DEBUG 819 printf("%s: Returning %d bytes.\n", keg->uk_name, 820 PAGE_SIZE * keg->uk_ppera); 821#endif 822 keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); 823} 824 825/* 826 * Frees pages from a keg back to the system. This is done on demand from 827 * the pageout daemon. 828 * 829 * Returns nothing. 830 */ 831static void 832keg_drain(uma_keg_t keg) 833{ 834 struct slabhead freeslabs = { 0 }; 835 uma_slab_t slab; 836 uma_slab_t n; 837 838 /* 839 * We don't want to take pages from statically allocated kegs at this 840 * time 841 */ 842 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) 843 return; 844 845#ifdef UMA_DEBUG 846 printf("%s free items: %u\n", keg->uk_name, keg->uk_free); 847#endif 848 KEG_LOCK(keg); 849 if (keg->uk_free == 0) 850 goto finished; 851 852 slab = LIST_FIRST(&keg->uk_free_slab); 853 while (slab) { 854 n = LIST_NEXT(slab, us_link); 855 856 /* We have no where to free these to */ 857 if (slab->us_flags & UMA_SLAB_BOOT) { 858 slab = n; 859 continue; 860 } 861 862 LIST_REMOVE(slab, us_link); 863 keg->uk_pages -= keg->uk_ppera; 864 keg->uk_free -= keg->uk_ipers; 865 866 if (keg->uk_flags & UMA_ZONE_HASH) 867 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); 868 869 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 870 871 slab = n; 872 } 873finished: 874 KEG_UNLOCK(keg); 875 876 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 877 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 878 keg_free_slab(keg, slab, keg->uk_ipers); 879 } 880} 881 882static void 883zone_drain_wait(uma_zone_t zone, int waitok) 884{ 885 886 /* 887 * Set draining to interlock with zone_dtor() so we can release our 888 * locks as we go. Only dtor() should do a WAITOK call since it 889 * is the only call that knows the structure will still be available 890 * when it wakes up. 891 */ 892 ZONE_LOCK(zone); 893 while (zone->uz_flags & UMA_ZFLAG_DRAINING) { 894 if (waitok == M_NOWAIT) 895 goto out; 896 mtx_unlock(&uma_mtx); 897 msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); 898 mtx_lock(&uma_mtx); 899 } 900 zone->uz_flags |= UMA_ZFLAG_DRAINING; 901 bucket_cache_drain(zone); 902 ZONE_UNLOCK(zone); 903 /* 904 * The DRAINING flag protects us from being freed while 905 * we're running. Normally the uma_mtx would protect us but we 906 * must be able to release and acquire the right lock for each keg. 907 */ 908 zone_foreach_keg(zone, &keg_drain); 909 ZONE_LOCK(zone); 910 zone->uz_flags &= ~UMA_ZFLAG_DRAINING; 911 wakeup(zone); 912out: 913 ZONE_UNLOCK(zone); 914} 915 916void 917zone_drain(uma_zone_t zone) 918{ 919 920 zone_drain_wait(zone, M_NOWAIT); 921} 922 923/* 924 * Allocate a new slab for a keg. This does not insert the slab onto a list. 925 * 926 * Arguments: 927 * wait Shall we wait? 928 * 929 * Returns: 930 * The slab that was allocated or NULL if there is no memory and the 931 * caller specified M_NOWAIT. 932 */ 933static uma_slab_t 934keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) 935{ 936 uma_slabrefcnt_t slabref; 937 uma_alloc allocf; 938 uma_slab_t slab; 939 uint8_t *mem; 940 uint8_t flags; 941 int i; 942 943 mtx_assert(&keg->uk_lock, MA_OWNED); 944 slab = NULL; 945 mem = NULL; 946 947#ifdef UMA_DEBUG 948 printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name); 949#endif 950 allocf = keg->uk_allocf; 951 KEG_UNLOCK(keg); 952 953 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 954 slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); 955 if (slab == NULL) 956 goto out; 957 } 958 959 /* 960 * This reproduces the old vm_zone behavior of zero filling pages the 961 * first time they are added to a zone. 962 * 963 * Malloced items are zeroed in uma_zalloc. 964 */ 965 966 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) 967 wait |= M_ZERO; 968 else 969 wait &= ~M_ZERO; 970 971 if (keg->uk_flags & UMA_ZONE_NODUMP) 972 wait |= M_NODUMP; 973 974 /* zone is passed for legacy reasons. */ 975 mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait); 976 if (mem == NULL) { 977 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 978 zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); 979 slab = NULL; 980 goto out; 981 } 982 983 /* Point the slab into the allocated memory */ 984 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) 985 slab = (uma_slab_t )(mem + keg->uk_pgoff); 986 987 if (keg->uk_flags & UMA_ZONE_VTOSLAB) 988 for (i = 0; i < keg->uk_ppera; i++) 989 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 990 991 slab->us_keg = keg; 992 slab->us_data = mem; 993 slab->us_freecount = keg->uk_ipers; 994 slab->us_flags = flags; 995 BIT_FILL(SLAB_SETSIZE, &slab->us_free); 996#ifdef INVARIANTS 997 BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); 998#endif 999 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1000 slabref = (uma_slabrefcnt_t)slab; 1001 for (i = 0; i < keg->uk_ipers; i++) 1002 slabref->us_refcnt[i] = 0; 1003 } 1004 1005 if (keg->uk_init != NULL) { 1006 for (i = 0; i < keg->uk_ipers; i++) 1007 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), 1008 keg->uk_size, wait) != 0) 1009 break; 1010 if (i != keg->uk_ipers) { 1011 keg_free_slab(keg, slab, i); 1012 slab = NULL; 1013 goto out; 1014 } 1015 } 1016out: 1017 KEG_LOCK(keg); 1018 1019 if (slab != NULL) { 1020 if (keg->uk_flags & UMA_ZONE_HASH) 1021 UMA_HASH_INSERT(&keg->uk_hash, slab, mem); 1022 1023 keg->uk_pages += keg->uk_ppera; 1024 keg->uk_free += keg->uk_ipers; 1025 } 1026 1027 return (slab); 1028} 1029 1030/* 1031 * This function is intended to be used early on in place of page_alloc() so 1032 * that we may use the boot time page cache to satisfy allocations before 1033 * the VM is ready. 1034 */ 1035static void * 1036startup_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait) 1037{ 1038 uma_keg_t keg; 1039 uma_slab_t tmps; 1040 int pages, check_pages; 1041 1042 keg = zone_first_keg(zone); 1043 pages = howmany(bytes, PAGE_SIZE); 1044 check_pages = pages - 1; 1045 KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n")); 1046 1047 /* 1048 * Check our small startup cache to see if it has pages remaining. 1049 */ 1050 mtx_lock(&uma_boot_pages_mtx); 1051 1052 /* First check if we have enough room. */ 1053 tmps = LIST_FIRST(&uma_boot_pages); 1054 while (tmps != NULL && check_pages-- > 0) 1055 tmps = LIST_NEXT(tmps, us_link); 1056 if (tmps != NULL) { 1057 /* 1058 * It's ok to lose tmps references. The last one will 1059 * have tmps->us_data pointing to the start address of 1060 * "pages" contiguous pages of memory. 1061 */ 1062 while (pages-- > 0) { 1063 tmps = LIST_FIRST(&uma_boot_pages); 1064 LIST_REMOVE(tmps, us_link); 1065 } 1066 mtx_unlock(&uma_boot_pages_mtx); 1067 *pflag = tmps->us_flags; 1068 return (tmps->us_data); 1069 } 1070 mtx_unlock(&uma_boot_pages_mtx); 1071 if (booted < UMA_STARTUP2) 1072 panic("UMA: Increase vm.boot_pages"); 1073 /* 1074 * Now that we've booted reset these users to their real allocator. 1075 */ 1076#ifdef UMA_MD_SMALL_ALLOC 1077 keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; 1078#else 1079 keg->uk_allocf = page_alloc; 1080#endif 1081 return keg->uk_allocf(zone, bytes, pflag, wait); 1082} 1083 1084/* 1085 * Allocates a number of pages from the system 1086 * 1087 * Arguments: 1088 * bytes The number of bytes requested 1089 * wait Shall we wait? 1090 * 1091 * Returns: 1092 * A pointer to the alloced memory or possibly 1093 * NULL if M_NOWAIT is set. 1094 */ 1095static void * 1096page_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait) 1097{ 1098 void *p; /* Returned page */ 1099 1100 *pflag = UMA_SLAB_KMEM; 1101 p = (void *) kmem_malloc(kmem_arena, bytes, wait); 1102 1103 return (p); 1104} 1105 1106/* 1107 * Allocates a number of pages from within an object 1108 * 1109 * Arguments: 1110 * bytes The number of bytes requested 1111 * wait Shall we wait? 1112 * 1113 * Returns: 1114 * A pointer to the alloced memory or possibly 1115 * NULL if M_NOWAIT is set. 1116 */ 1117static void * 1118noobj_alloc(uma_zone_t zone, int bytes, uint8_t *flags, int wait) 1119{ 1120 TAILQ_HEAD(, vm_page) alloctail; 1121 u_long npages; 1122 vm_offset_t retkva, zkva; 1123 vm_page_t p, p_next; 1124 uma_keg_t keg; 1125 1126 TAILQ_INIT(&alloctail); 1127 keg = zone_first_keg(zone); 1128 1129 npages = howmany(bytes, PAGE_SIZE); 1130 while (npages > 0) { 1131 p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT | 1132 VM_ALLOC_WIRED | VM_ALLOC_NOOBJ); 1133 if (p != NULL) { 1134 /* 1135 * Since the page does not belong to an object, its 1136 * listq is unused. 1137 */ 1138 TAILQ_INSERT_TAIL(&alloctail, p, listq); 1139 npages--; 1140 continue; 1141 } 1142 if (wait & M_WAITOK) { 1143 VM_WAIT; 1144 continue; 1145 } 1146 1147 /* 1148 * Page allocation failed, free intermediate pages and 1149 * exit. 1150 */ 1151 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { 1152 vm_page_unwire(p, 0); 1153 vm_page_free(p); 1154 } 1155 return (NULL); 1156 } 1157 *flags = UMA_SLAB_PRIV; 1158 zkva = keg->uk_kva + 1159 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); 1160 retkva = zkva; 1161 TAILQ_FOREACH(p, &alloctail, listq) { 1162 pmap_qenter(zkva, &p, 1); 1163 zkva += PAGE_SIZE; 1164 } 1165 1166 return ((void *)retkva); 1167} 1168 1169/* 1170 * Frees a number of pages to the system 1171 * 1172 * Arguments: 1173 * mem A pointer to the memory to be freed 1174 * size The size of the memory being freed 1175 * flags The original p->us_flags field 1176 * 1177 * Returns: 1178 * Nothing 1179 */ 1180static void 1181page_free(void *mem, int size, uint8_t flags) 1182{ 1183 struct vmem *vmem; 1184 1185 if (flags & UMA_SLAB_KMEM) 1186 vmem = kmem_arena; 1187 else if (flags & UMA_SLAB_KERNEL) 1188 vmem = kernel_arena; 1189 else 1190 panic("UMA: page_free used with invalid flags %d", flags); 1191 1192 kmem_free(vmem, (vm_offset_t)mem, size); 1193} 1194 1195/* 1196 * Zero fill initializer 1197 * 1198 * Arguments/Returns follow uma_init specifications 1199 */ 1200static int 1201zero_init(void *mem, int size, int flags) 1202{ 1203 bzero(mem, size); 1204 return (0); 1205} 1206 1207/* 1208 * Finish creating a small uma keg. This calculates ipers, and the keg size. 1209 * 1210 * Arguments 1211 * keg The zone we should initialize 1212 * 1213 * Returns 1214 * Nothing 1215 */ 1216static void 1217keg_small_init(uma_keg_t keg) 1218{ 1219 u_int rsize; 1220 u_int memused; 1221 u_int wastedspace; 1222 u_int shsize; 1223 1224 if (keg->uk_flags & UMA_ZONE_PCPU) { 1225 u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU; 1226 1227 keg->uk_slabsize = sizeof(struct pcpu); 1228 keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu), 1229 PAGE_SIZE); 1230 } else { 1231 keg->uk_slabsize = UMA_SLAB_SIZE; 1232 keg->uk_ppera = 1; 1233 } 1234 1235 /* 1236 * Calculate the size of each allocation (rsize) according to 1237 * alignment. If the requested size is smaller than we have 1238 * allocation bits for we round it up. 1239 */ 1240 rsize = keg->uk_size; 1241 if (rsize < keg->uk_slabsize / SLAB_SETSIZE) 1242 rsize = keg->uk_slabsize / SLAB_SETSIZE; 1243 if (rsize & keg->uk_align) 1244 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); 1245 keg->uk_rsize = rsize; 1246 1247 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || 1248 keg->uk_rsize < sizeof(struct pcpu), 1249 ("%s: size %u too large", __func__, keg->uk_rsize)); 1250 1251 if (keg->uk_flags & UMA_ZONE_REFCNT) 1252 rsize += sizeof(uint32_t); 1253 1254 if (keg->uk_flags & UMA_ZONE_OFFPAGE) 1255 shsize = 0; 1256 else 1257 shsize = sizeof(struct uma_slab); 1258 1259 keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize; 1260 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1261 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1262 1263 memused = keg->uk_ipers * rsize + shsize; 1264 wastedspace = keg->uk_slabsize - memused; 1265 1266 /* 1267 * We can't do OFFPAGE if we're internal or if we've been 1268 * asked to not go to the VM for buckets. If we do this we 1269 * may end up going to the VM for slabs which we do not 1270 * want to do if we're UMA_ZFLAG_CACHEONLY as a result 1271 * of UMA_ZONE_VM, which clearly forbids it. 1272 */ 1273 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || 1274 (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) 1275 return; 1276 1277 /* 1278 * See if using an OFFPAGE slab will limit our waste. Only do 1279 * this if it permits more items per-slab. 1280 * 1281 * XXX We could try growing slabsize to limit max waste as well. 1282 * Historically this was not done because the VM could not 1283 * efficiently handle contiguous allocations. 1284 */ 1285 if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) && 1286 (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) { 1287 keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize; 1288 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, 1289 ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); 1290#ifdef UMA_DEBUG 1291 printf("UMA decided we need offpage slab headers for " 1292 "keg: %s, calculated wastedspace = %d, " 1293 "maximum wasted space allowed = %d, " 1294 "calculated ipers = %d, " 1295 "new wasted space = %d\n", keg->uk_name, wastedspace, 1296 keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers, 1297 keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize); 1298#endif 1299 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1300 } 1301 1302 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1303 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1304 keg->uk_flags |= UMA_ZONE_HASH; 1305} 1306 1307/* 1308 * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do 1309 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 1310 * more complicated. 1311 * 1312 * Arguments 1313 * keg The keg we should initialize 1314 * 1315 * Returns 1316 * Nothing 1317 */ 1318static void 1319keg_large_init(uma_keg_t keg) 1320{ 1321 u_int shsize; 1322 1323 KASSERT(keg != NULL, ("Keg is null in keg_large_init")); 1324 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, 1325 ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); 1326 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1327 ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); 1328 1329 keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); 1330 keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE; 1331 keg->uk_ipers = 1; 1332 keg->uk_rsize = keg->uk_size; 1333 1334 /* We can't do OFFPAGE if we're internal, bail out here. */ 1335 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) 1336 return; 1337 1338 /* Check whether we have enough space to not do OFFPAGE. */ 1339 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) { 1340 shsize = sizeof(struct uma_slab); 1341 if (keg->uk_flags & UMA_ZONE_REFCNT) 1342 shsize += keg->uk_ipers * sizeof(uint32_t); 1343 if (shsize & UMA_ALIGN_PTR) 1344 shsize = (shsize & ~UMA_ALIGN_PTR) + 1345 (UMA_ALIGN_PTR + 1); 1346 1347 if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize) 1348 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1349 } 1350 1351 if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && 1352 (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) 1353 keg->uk_flags |= UMA_ZONE_HASH; 1354} 1355 1356static void 1357keg_cachespread_init(uma_keg_t keg) 1358{ 1359 int alignsize; 1360 int trailer; 1361 int pages; 1362 int rsize; 1363 1364 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, 1365 ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); 1366 1367 alignsize = keg->uk_align + 1; 1368 rsize = keg->uk_size; 1369 /* 1370 * We want one item to start on every align boundary in a page. To 1371 * do this we will span pages. We will also extend the item by the 1372 * size of align if it is an even multiple of align. Otherwise, it 1373 * would fall on the same boundary every time. 1374 */ 1375 if (rsize & keg->uk_align) 1376 rsize = (rsize & ~keg->uk_align) + alignsize; 1377 if ((rsize & alignsize) == 0) 1378 rsize += alignsize; 1379 trailer = rsize - keg->uk_size; 1380 pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; 1381 pages = MIN(pages, (128 * 1024) / PAGE_SIZE); 1382 keg->uk_rsize = rsize; 1383 keg->uk_ppera = pages; 1384 keg->uk_slabsize = UMA_SLAB_SIZE; 1385 keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; 1386 keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; 1387 KASSERT(keg->uk_ipers <= uma_max_ipers, 1388 ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, 1389 keg->uk_ipers)); 1390} 1391 1392/* 1393 * Keg header ctor. This initializes all fields, locks, etc. And inserts 1394 * the keg onto the global keg list. 1395 * 1396 * Arguments/Returns follow uma_ctor specifications 1397 * udata Actually uma_kctor_args 1398 */ 1399static int 1400keg_ctor(void *mem, int size, void *udata, int flags) 1401{ 1402 struct uma_kctor_args *arg = udata; 1403 uma_keg_t keg = mem; 1404 uma_zone_t zone; 1405 1406 bzero(keg, size); 1407 keg->uk_size = arg->size; 1408 keg->uk_init = arg->uminit; 1409 keg->uk_fini = arg->fini; 1410 keg->uk_align = arg->align; 1411 keg->uk_free = 0; 1412 keg->uk_reserve = 0; 1413 keg->uk_pages = 0; 1414 keg->uk_flags = arg->flags; 1415 keg->uk_allocf = page_alloc; 1416 keg->uk_freef = page_free; 1417 keg->uk_slabzone = NULL; 1418 1419 /* 1420 * The master zone is passed to us at keg-creation time. 1421 */ 1422 zone = arg->zone; 1423 keg->uk_name = zone->uz_name; 1424 1425 if (arg->flags & UMA_ZONE_VM) 1426 keg->uk_flags |= UMA_ZFLAG_CACHEONLY; 1427 1428 if (arg->flags & UMA_ZONE_ZINIT) 1429 keg->uk_init = zero_init; 1430 1431 if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC) 1432 keg->uk_flags |= UMA_ZONE_VTOSLAB; 1433 1434 if (arg->flags & UMA_ZONE_PCPU) 1435#ifdef SMP 1436 keg->uk_flags |= UMA_ZONE_OFFPAGE; 1437#else 1438 keg->uk_flags &= ~UMA_ZONE_PCPU; 1439#endif 1440 1441 if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { 1442 keg_cachespread_init(keg); 1443 } else if (keg->uk_flags & UMA_ZONE_REFCNT) { 1444 if (keg->uk_size > 1445 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - 1446 sizeof(uint32_t))) 1447 keg_large_init(keg); 1448 else 1449 keg_small_init(keg); 1450 } else { 1451 if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab))) 1452 keg_large_init(keg); 1453 else 1454 keg_small_init(keg); 1455 } 1456 1457 if (keg->uk_flags & UMA_ZONE_OFFPAGE) { 1458 if (keg->uk_flags & UMA_ZONE_REFCNT) { 1459 if (keg->uk_ipers > uma_max_ipers_ref) 1460 panic("Too many ref items per zone: %d > %d\n", 1461 keg->uk_ipers, uma_max_ipers_ref); 1462 keg->uk_slabzone = slabrefzone; 1463 } else 1464 keg->uk_slabzone = slabzone; 1465 } 1466 1467 /* 1468 * If we haven't booted yet we need allocations to go through the 1469 * startup cache until the vm is ready. 1470 */ 1471 if (keg->uk_ppera == 1) { 1472#ifdef UMA_MD_SMALL_ALLOC 1473 keg->uk_allocf = uma_small_alloc; 1474 keg->uk_freef = uma_small_free; 1475 1476 if (booted < UMA_STARTUP) 1477 keg->uk_allocf = startup_alloc; 1478#else 1479 if (booted < UMA_STARTUP2) 1480 keg->uk_allocf = startup_alloc; 1481#endif 1482 } else if (booted < UMA_STARTUP2 && 1483 (keg->uk_flags & UMA_ZFLAG_INTERNAL)) 1484 keg->uk_allocf = startup_alloc; 1485 1486 /* 1487 * Initialize keg's lock 1488 */ 1489 KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); 1490 1491 /* 1492 * If we're putting the slab header in the actual page we need to 1493 * figure out where in each page it goes. This calculates a right 1494 * justified offset into the memory on an ALIGN_PTR boundary. 1495 */ 1496 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { 1497 u_int totsize; 1498 1499 /* Size of the slab struct and free list */ 1500 totsize = sizeof(struct uma_slab); 1501 1502 /* Size of the reference counts. */ 1503 if (keg->uk_flags & UMA_ZONE_REFCNT) 1504 totsize += keg->uk_ipers * sizeof(uint32_t); 1505 1506 if (totsize & UMA_ALIGN_PTR) 1507 totsize = (totsize & ~UMA_ALIGN_PTR) + 1508 (UMA_ALIGN_PTR + 1); 1509 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; 1510 1511 /* 1512 * The only way the following is possible is if with our 1513 * UMA_ALIGN_PTR adjustments we are now bigger than 1514 * UMA_SLAB_SIZE. I haven't checked whether this is 1515 * mathematically possible for all cases, so we make 1516 * sure here anyway. 1517 */ 1518 totsize = keg->uk_pgoff + sizeof(struct uma_slab); 1519 if (keg->uk_flags & UMA_ZONE_REFCNT) 1520 totsize += keg->uk_ipers * sizeof(uint32_t); 1521 if (totsize > PAGE_SIZE * keg->uk_ppera) { 1522 printf("zone %s ipers %d rsize %d size %d\n", 1523 zone->uz_name, keg->uk_ipers, keg->uk_rsize, 1524 keg->uk_size); 1525 panic("UMA slab won't fit."); 1526 } 1527 } 1528 1529 if (keg->uk_flags & UMA_ZONE_HASH) 1530 hash_alloc(&keg->uk_hash); 1531 1532#ifdef UMA_DEBUG 1533 printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n", 1534 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags, 1535 keg->uk_ipers, keg->uk_ppera, 1536 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free); 1537#endif 1538 1539 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); 1540 1541 mtx_lock(&uma_mtx); 1542 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); 1543 mtx_unlock(&uma_mtx); 1544 return (0); 1545} 1546 1547/* 1548 * Zone header ctor. This initializes all fields, locks, etc. 1549 * 1550 * Arguments/Returns follow uma_ctor specifications 1551 * udata Actually uma_zctor_args 1552 */ 1553static int 1554zone_ctor(void *mem, int size, void *udata, int flags) 1555{ 1556 struct uma_zctor_args *arg = udata; 1557 uma_zone_t zone = mem; 1558 uma_zone_t z; 1559 uma_keg_t keg; 1560 1561 bzero(zone, size); 1562 zone->uz_name = arg->name; 1563 zone->uz_ctor = arg->ctor; 1564 zone->uz_dtor = arg->dtor; 1565 zone->uz_slab = zone_fetch_slab; 1566 zone->uz_init = NULL; 1567 zone->uz_fini = NULL; 1568 zone->uz_allocs = 0; 1569 zone->uz_frees = 0; 1570 zone->uz_fails = 0; 1571 zone->uz_sleeps = 0; 1572 zone->uz_count = 0; 1573 zone->uz_count_min = 0; 1574 zone->uz_flags = 0; 1575 zone->uz_warning = NULL; 1576 timevalclear(&zone->uz_ratecheck); 1577 keg = arg->keg; 1578 1579 ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); 1580 1581 /* 1582 * This is a pure cache zone, no kegs. 1583 */ 1584 if (arg->import) { 1585 if (arg->flags & UMA_ZONE_VM) 1586 arg->flags |= UMA_ZFLAG_CACHEONLY; 1587 zone->uz_flags = arg->flags; 1588 zone->uz_size = arg->size; 1589 zone->uz_import = arg->import; 1590 zone->uz_release = arg->release; 1591 zone->uz_arg = arg->arg; 1592 zone->uz_lockptr = &zone->uz_lock; 1593 goto out; 1594 } 1595 1596 /* 1597 * Use the regular zone/keg/slab allocator. 1598 */ 1599 zone->uz_import = (uma_import)zone_import; 1600 zone->uz_release = (uma_release)zone_release; 1601 zone->uz_arg = zone; 1602 1603 if (arg->flags & UMA_ZONE_SECONDARY) { 1604 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); 1605 zone->uz_init = arg->uminit; 1606 zone->uz_fini = arg->fini; 1607 zone->uz_lockptr = &keg->uk_lock; 1608 zone->uz_flags |= UMA_ZONE_SECONDARY; 1609 mtx_lock(&uma_mtx); 1610 ZONE_LOCK(zone); 1611 LIST_FOREACH(z, &keg->uk_zones, uz_link) { 1612 if (LIST_NEXT(z, uz_link) == NULL) { 1613 LIST_INSERT_AFTER(z, zone, uz_link); 1614 break; 1615 } 1616 } 1617 ZONE_UNLOCK(zone); 1618 mtx_unlock(&uma_mtx); 1619 } else if (keg == NULL) { 1620 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, 1621 arg->align, arg->flags)) == NULL) 1622 return (ENOMEM); 1623 } else { 1624 struct uma_kctor_args karg; 1625 int error; 1626 1627 /* We should only be here from uma_startup() */ 1628 karg.size = arg->size; 1629 karg.uminit = arg->uminit; 1630 karg.fini = arg->fini; 1631 karg.align = arg->align; 1632 karg.flags = arg->flags; 1633 karg.zone = zone; 1634 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, 1635 flags); 1636 if (error) 1637 return (error); 1638 } 1639 1640 /* 1641 * Link in the first keg. 1642 */ 1643 zone->uz_klink.kl_keg = keg; 1644 LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); 1645 zone->uz_lockptr = &keg->uk_lock; 1646 zone->uz_size = keg->uk_size; 1647 zone->uz_flags |= (keg->uk_flags & 1648 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); 1649 1650 /* 1651 * Some internal zones don't have room allocated for the per cpu 1652 * caches. If we're internal, bail out here. 1653 */ 1654 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { 1655 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, 1656 ("Secondary zone requested UMA_ZFLAG_INTERNAL")); 1657 return (0); 1658 } 1659 1660out: 1661 if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0) 1662 zone->uz_count = bucket_select(zone->uz_size); 1663 else 1664 zone->uz_count = BUCKET_MAX; 1665 zone->uz_count_min = zone->uz_count; 1666 1667 return (0); 1668} 1669 1670/* 1671 * Keg header dtor. This frees all data, destroys locks, frees the hash 1672 * table and removes the keg from the global list. 1673 * 1674 * Arguments/Returns follow uma_dtor specifications 1675 * udata unused 1676 */ 1677static void 1678keg_dtor(void *arg, int size, void *udata) 1679{ 1680 uma_keg_t keg; 1681 1682 keg = (uma_keg_t)arg; 1683 KEG_LOCK(keg); 1684 if (keg->uk_free != 0) { 1685 printf("Freed UMA keg (%s) was not empty (%d items). " 1686 " Lost %d pages of memory.\n", 1687 keg->uk_name ? keg->uk_name : "", 1688 keg->uk_free, keg->uk_pages); 1689 } 1690 KEG_UNLOCK(keg); 1691 1692 hash_free(&keg->uk_hash); 1693 1694 KEG_LOCK_FINI(keg); 1695} 1696 1697/* 1698 * Zone header dtor. 1699 * 1700 * Arguments/Returns follow uma_dtor specifications 1701 * udata unused 1702 */ 1703static void 1704zone_dtor(void *arg, int size, void *udata) 1705{ 1706 uma_klink_t klink; 1707 uma_zone_t zone; 1708 uma_keg_t keg; 1709 1710 zone = (uma_zone_t)arg; 1711 keg = zone_first_keg(zone); 1712 1713 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 1714 cache_drain(zone); 1715 1716 mtx_lock(&uma_mtx); 1717 LIST_REMOVE(zone, uz_link); 1718 mtx_unlock(&uma_mtx); 1719 /* 1720 * XXX there are some races here where 1721 * the zone can be drained but zone lock 1722 * released and then refilled before we 1723 * remove it... we dont care for now 1724 */ 1725 zone_drain_wait(zone, M_WAITOK); 1726 /* 1727 * Unlink all of our kegs. 1728 */ 1729 while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { 1730 klink->kl_keg = NULL; 1731 LIST_REMOVE(klink, kl_link); 1732 if (klink == &zone->uz_klink) 1733 continue; 1734 free(klink, M_TEMP); 1735 } 1736 /* 1737 * We only destroy kegs from non secondary zones. 1738 */ 1739 if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { 1740 mtx_lock(&uma_mtx); 1741 LIST_REMOVE(keg, uk_link); 1742 mtx_unlock(&uma_mtx); 1743 zone_free_item(kegs, keg, NULL, SKIP_NONE); 1744 } 1745 ZONE_LOCK_FINI(zone); 1746} 1747 1748/* 1749 * Traverses every zone in the system and calls a callback 1750 * 1751 * Arguments: 1752 * zfunc A pointer to a function which accepts a zone 1753 * as an argument. 1754 * 1755 * Returns: 1756 * Nothing 1757 */ 1758static void 1759zone_foreach(void (*zfunc)(uma_zone_t)) 1760{ 1761 uma_keg_t keg; 1762 uma_zone_t zone; 1763 1764 mtx_lock(&uma_mtx); 1765 LIST_FOREACH(keg, &uma_kegs, uk_link) { 1766 LIST_FOREACH(zone, &keg->uk_zones, uz_link) 1767 zfunc(zone); 1768 } 1769 mtx_unlock(&uma_mtx); 1770} 1771 1772/* Public functions */ 1773/* See uma.h */ 1774void 1775uma_startup(void *bootmem, int boot_pages) 1776{ 1777 struct uma_zctor_args args; 1778 uma_slab_t slab; 1779 u_int slabsize; 1780 int i; 1781 1782#ifdef UMA_DEBUG 1783 printf("Creating uma keg headers zone and keg.\n"); 1784#endif 1785 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF); 1786 1787 /* "manually" create the initial zone */ 1788 memset(&args, 0, sizeof(args)); 1789 args.name = "UMA Kegs"; 1790 args.size = sizeof(struct uma_keg); 1791 args.ctor = keg_ctor; 1792 args.dtor = keg_dtor; 1793 args.uminit = zero_init; 1794 args.fini = NULL; 1795 args.keg = &masterkeg; 1796 args.align = 32 - 1; 1797 args.flags = UMA_ZFLAG_INTERNAL; 1798 /* The initial zone has no Per cpu queues so it's smaller */ 1799 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); 1800 1801#ifdef UMA_DEBUG 1802 printf("Filling boot free list.\n"); 1803#endif 1804 for (i = 0; i < boot_pages; i++) { 1805 slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE)); 1806 slab->us_data = (uint8_t *)slab; 1807 slab->us_flags = UMA_SLAB_BOOT; 1808 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); 1809 } 1810 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); 1811 1812#ifdef UMA_DEBUG 1813 printf("Creating uma zone headers zone and keg.\n"); 1814#endif 1815 args.name = "UMA Zones"; 1816 args.size = sizeof(struct uma_zone) + 1817 (sizeof(struct uma_cache) * (mp_maxid + 1)); 1818 args.ctor = zone_ctor; 1819 args.dtor = zone_dtor; 1820 args.uminit = zero_init; 1821 args.fini = NULL; 1822 args.keg = NULL; 1823 args.align = 32 - 1; 1824 args.flags = UMA_ZFLAG_INTERNAL; 1825 /* The initial zone has no Per cpu queues so it's smaller */ 1826 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); 1827 1828#ifdef UMA_DEBUG 1829 printf("Initializing pcpu cache locks.\n"); 1830#endif 1831#ifdef UMA_DEBUG 1832 printf("Creating slab and hash zones.\n"); 1833#endif 1834 1835 /* Now make a zone for slab headers */ 1836 slabzone = uma_zcreate("UMA Slabs", 1837 sizeof(struct uma_slab), 1838 NULL, NULL, NULL, NULL, 1839 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1840 1841 /* 1842 * We also create a zone for the bigger slabs with reference 1843 * counts in them, to accomodate UMA_ZONE_REFCNT zones. 1844 */ 1845 slabsize = sizeof(struct uma_slab_refcnt); 1846 slabsize += uma_max_ipers_ref * sizeof(uint32_t); 1847 slabrefzone = uma_zcreate("UMA RCntSlabs", 1848 slabsize, 1849 NULL, NULL, NULL, NULL, 1850 UMA_ALIGN_PTR, 1851 UMA_ZFLAG_INTERNAL); 1852 1853 hashzone = uma_zcreate("UMA Hash", 1854 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 1855 NULL, NULL, NULL, NULL, 1856 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); 1857 1858 bucket_init(); 1859 1860 booted = UMA_STARTUP; 1861 1862#ifdef UMA_DEBUG 1863 printf("UMA startup complete.\n"); 1864#endif 1865} 1866 1867/* see uma.h */ 1868void 1869uma_startup2(void) 1870{ 1871 booted = UMA_STARTUP2; 1872 bucket_enable(); 1873#ifdef UMA_DEBUG 1874 printf("UMA startup2 complete.\n"); 1875#endif 1876} 1877 1878/* 1879 * Initialize our callout handle 1880 * 1881 */ 1882 1883static void 1884uma_startup3(void) 1885{ 1886#ifdef UMA_DEBUG 1887 printf("Starting callout.\n"); 1888#endif 1889 callout_init(&uma_callout, CALLOUT_MPSAFE); 1890 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); 1891#ifdef UMA_DEBUG 1892 printf("UMA startup3 complete.\n"); 1893#endif 1894} 1895 1896static uma_keg_t 1897uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, 1898 int align, uint32_t flags) 1899{ 1900 struct uma_kctor_args args; 1901 1902 args.size = size; 1903 args.uminit = uminit; 1904 args.fini = fini; 1905 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; 1906 args.flags = flags; 1907 args.zone = zone; 1908 return (zone_alloc_item(kegs, &args, M_WAITOK)); 1909} 1910 1911/* See uma.h */ 1912void 1913uma_set_align(int align) 1914{ 1915 1916 if (align != UMA_ALIGN_CACHE) 1917 uma_align_cache = align; 1918} 1919 1920/* See uma.h */ 1921uma_zone_t 1922uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 1923 uma_init uminit, uma_fini fini, int align, uint32_t flags) 1924 1925{ 1926 struct uma_zctor_args args; 1927 1928 /* This stuff is essential for the zone ctor */ 1929 memset(&args, 0, sizeof(args)); 1930 args.name = name; 1931 args.size = size; 1932 args.ctor = ctor; 1933 args.dtor = dtor; 1934 args.uminit = uminit; 1935 args.fini = fini; 1936 args.align = align; 1937 args.flags = flags; 1938 args.keg = NULL; 1939 1940 return (zone_alloc_item(zones, &args, M_WAITOK)); 1941} 1942 1943/* See uma.h */ 1944uma_zone_t 1945uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, 1946 uma_init zinit, uma_fini zfini, uma_zone_t master) 1947{ 1948 struct uma_zctor_args args; 1949 uma_keg_t keg; 1950 1951 keg = zone_first_keg(master); 1952 memset(&args, 0, sizeof(args)); 1953 args.name = name; 1954 args.size = keg->uk_size; 1955 args.ctor = ctor; 1956 args.dtor = dtor; 1957 args.uminit = zinit; 1958 args.fini = zfini; 1959 args.align = keg->uk_align; 1960 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; 1961 args.keg = keg; 1962 1963 /* XXX Attaches only one keg of potentially many. */ 1964 return (zone_alloc_item(zones, &args, M_WAITOK)); 1965} 1966 1967/* See uma.h */ 1968uma_zone_t 1969uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, 1970 uma_init zinit, uma_fini zfini, uma_import zimport, 1971 uma_release zrelease, void *arg, int flags) 1972{ 1973 struct uma_zctor_args args; 1974 1975 memset(&args, 0, sizeof(args)); 1976 args.name = name; 1977 args.size = size; 1978 args.ctor = ctor; 1979 args.dtor = dtor; 1980 args.uminit = zinit; 1981 args.fini = zfini; 1982 args.import = zimport; 1983 args.release = zrelease; 1984 args.arg = arg; 1985 args.align = 0; 1986 args.flags = flags; 1987 1988 return (zone_alloc_item(zones, &args, M_WAITOK)); 1989} 1990 1991static void 1992zone_lock_pair(uma_zone_t a, uma_zone_t b) 1993{ 1994 if (a < b) { 1995 ZONE_LOCK(a); 1996 mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); 1997 } else { 1998 ZONE_LOCK(b); 1999 mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); 2000 } 2001} 2002 2003static void 2004zone_unlock_pair(uma_zone_t a, uma_zone_t b) 2005{ 2006 2007 ZONE_UNLOCK(a); 2008 ZONE_UNLOCK(b); 2009} 2010 2011int 2012uma_zsecond_add(uma_zone_t zone, uma_zone_t master) 2013{ 2014 uma_klink_t klink; 2015 uma_klink_t kl; 2016 int error; 2017 2018 error = 0; 2019 klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); 2020 2021 zone_lock_pair(zone, master); 2022 /* 2023 * zone must use vtoslab() to resolve objects and must already be 2024 * a secondary. 2025 */ 2026 if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) 2027 != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { 2028 error = EINVAL; 2029 goto out; 2030 } 2031 /* 2032 * The new master must also use vtoslab(). 2033 */ 2034 if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { 2035 error = EINVAL; 2036 goto out; 2037 } 2038 /* 2039 * Both must either be refcnt, or not be refcnt. 2040 */ 2041 if ((zone->uz_flags & UMA_ZONE_REFCNT) != 2042 (master->uz_flags & UMA_ZONE_REFCNT)) { 2043 error = EINVAL; 2044 goto out; 2045 } 2046 /* 2047 * The underlying object must be the same size. rsize 2048 * may be different. 2049 */ 2050 if (master->uz_size != zone->uz_size) { 2051 error = E2BIG; 2052 goto out; 2053 } 2054 /* 2055 * Put it at the end of the list. 2056 */ 2057 klink->kl_keg = zone_first_keg(master); 2058 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { 2059 if (LIST_NEXT(kl, kl_link) == NULL) { 2060 LIST_INSERT_AFTER(kl, klink, kl_link); 2061 break; 2062 } 2063 } 2064 klink = NULL; 2065 zone->uz_flags |= UMA_ZFLAG_MULTI; 2066 zone->uz_slab = zone_fetch_slab_multi; 2067 2068out: 2069 zone_unlock_pair(zone, master); 2070 if (klink != NULL) 2071 free(klink, M_TEMP); 2072 2073 return (error); 2074} 2075 2076 2077/* See uma.h */ 2078void 2079uma_zdestroy(uma_zone_t zone) 2080{ 2081 2082 zone_free_item(zones, zone, NULL, SKIP_NONE); 2083} 2084 2085/* See uma.h */ 2086void * 2087uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 2088{ 2089 void *item; 2090 uma_cache_t cache; 2091 uma_bucket_t bucket; 2092 int lockfail; 2093 int cpu; 2094 2095 /* This is the fast path allocation */ 2096#ifdef UMA_DEBUG_ALLOC_1 2097 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); 2098#endif 2099 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread, 2100 zone->uz_name, flags); 2101 2102 if (flags & M_WAITOK) { 2103 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2104 "uma_zalloc_arg: zone \"%s\"", zone->uz_name); 2105 } 2106#ifdef DEBUG_MEMGUARD 2107 if (memguard_cmp_zone(zone)) { 2108 item = memguard_alloc(zone->uz_size, flags); 2109 if (item != NULL) { 2110 /* 2111 * Avoid conflict with the use-after-free 2112 * protecting infrastructure from INVARIANTS. 2113 */ 2114 if (zone->uz_init != NULL && 2115 zone->uz_init != mtrash_init && 2116 zone->uz_init(item, zone->uz_size, flags) != 0) 2117 return (NULL); 2118 if (zone->uz_ctor != NULL && 2119 zone->uz_ctor != mtrash_ctor && 2120 zone->uz_ctor(item, zone->uz_size, udata, 2121 flags) != 0) { 2122 zone->uz_fini(item, zone->uz_size); 2123 return (NULL); 2124 } 2125 return (item); 2126 } 2127 /* This is unfortunate but should not be fatal. */ 2128 } 2129#endif 2130 /* 2131 * If possible, allocate from the per-CPU cache. There are two 2132 * requirements for safe access to the per-CPU cache: (1) the thread 2133 * accessing the cache must not be preempted or yield during access, 2134 * and (2) the thread must not migrate CPUs without switching which 2135 * cache it accesses. We rely on a critical section to prevent 2136 * preemption and migration. We release the critical section in 2137 * order to acquire the zone mutex if we are unable to allocate from 2138 * the current cache; when we re-acquire the critical section, we 2139 * must detect and handle migration if it has occurred. 2140 */ 2141 critical_enter(); 2142 cpu = curcpu; 2143 cache = &zone->uz_cpu[cpu]; 2144 2145zalloc_start: 2146 bucket = cache->uc_allocbucket; 2147 if (bucket != NULL && bucket->ub_cnt > 0) { 2148 bucket->ub_cnt--; 2149 item = bucket->ub_bucket[bucket->ub_cnt]; 2150#ifdef INVARIANTS 2151 bucket->ub_bucket[bucket->ub_cnt] = NULL; 2152#endif 2153 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); 2154 cache->uc_allocs++; 2155 critical_exit(); 2156 if (zone->uz_ctor != NULL && 2157 zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2158 atomic_add_long(&zone->uz_fails, 1); 2159 zone_free_item(zone, item, udata, SKIP_DTOR); 2160 return (NULL); 2161 } 2162#ifdef INVARIANTS 2163 uma_dbg_alloc(zone, NULL, item); 2164#endif 2165 if (flags & M_ZERO) 2166 bzero(item, zone->uz_size); 2167 return (item); 2168 } 2169 2170 /* 2171 * We have run out of items in our alloc bucket. 2172 * See if we can switch with our free bucket. 2173 */ 2174 bucket = cache->uc_freebucket; 2175 if (bucket != NULL && bucket->ub_cnt > 0) { 2176#ifdef UMA_DEBUG_ALLOC 2177 printf("uma_zalloc: Swapping empty with alloc.\n"); 2178#endif 2179 cache->uc_freebucket = cache->uc_allocbucket; 2180 cache->uc_allocbucket = bucket; 2181 goto zalloc_start; 2182 } 2183 2184 /* 2185 * Discard any empty allocation bucket while we hold no locks. 2186 */ 2187 bucket = cache->uc_allocbucket; 2188 cache->uc_allocbucket = NULL; 2189 critical_exit(); 2190 if (bucket != NULL) 2191 bucket_free(zone, bucket, udata); 2192 2193 /* Short-circuit for zones without buckets and low memory. */ 2194 if (zone->uz_count == 0 || bucketdisable) 2195 goto zalloc_item; 2196 2197 /* 2198 * Attempt to retrieve the item from the per-CPU cache has failed, so 2199 * we must go back to the zone. This requires the zone lock, so we 2200 * must drop the critical section, then re-acquire it when we go back 2201 * to the cache. Since the critical section is released, we may be 2202 * preempted or migrate. As such, make sure not to maintain any 2203 * thread-local state specific to the cache from prior to releasing 2204 * the critical section. 2205 */ 2206 lockfail = 0; 2207 if (ZONE_TRYLOCK(zone) == 0) { 2208 /* Record contention to size the buckets. */ 2209 ZONE_LOCK(zone); 2210 lockfail = 1; 2211 } 2212 critical_enter(); 2213 cpu = curcpu; 2214 cache = &zone->uz_cpu[cpu]; 2215 2216 /* 2217 * Since we have locked the zone we may as well send back our stats. 2218 */ 2219 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2220 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2221 cache->uc_allocs = 0; 2222 cache->uc_frees = 0; 2223 2224 /* See if we lost the race to fill the cache. */ 2225 if (cache->uc_allocbucket != NULL) { 2226 ZONE_UNLOCK(zone); 2227 goto zalloc_start; 2228 } 2229 2230 /* 2231 * Check the zone's cache of buckets. 2232 */ 2233 if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { 2234 KASSERT(bucket->ub_cnt != 0, 2235 ("uma_zalloc_arg: Returning an empty bucket.")); 2236 2237 LIST_REMOVE(bucket, ub_link); 2238 cache->uc_allocbucket = bucket; 2239 ZONE_UNLOCK(zone); 2240 goto zalloc_start; 2241 } 2242 /* We are no longer associated with this CPU. */ 2243 critical_exit(); 2244 2245 /* 2246 * We bump the uz count when the cache size is insufficient to 2247 * handle the working set. 2248 */ 2249 if (lockfail && zone->uz_count < BUCKET_MAX) 2250 zone->uz_count++; 2251 ZONE_UNLOCK(zone); 2252 2253 /* 2254 * Now lets just fill a bucket and put it on the free list. If that 2255 * works we'll restart the allocation from the begining and it 2256 * will use the just filled bucket. 2257 */ 2258 bucket = zone_alloc_bucket(zone, udata, flags); 2259 if (bucket != NULL) { 2260 ZONE_LOCK(zone); 2261 critical_enter(); 2262 cpu = curcpu; 2263 cache = &zone->uz_cpu[cpu]; 2264 /* 2265 * See if we lost the race or were migrated. Cache the 2266 * initialized bucket to make this less likely or claim 2267 * the memory directly. 2268 */ 2269 if (cache->uc_allocbucket == NULL) 2270 cache->uc_allocbucket = bucket; 2271 else 2272 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2273 ZONE_UNLOCK(zone); 2274 goto zalloc_start; 2275 } 2276 2277 /* 2278 * We may not be able to get a bucket so return an actual item. 2279 */ 2280#ifdef UMA_DEBUG 2281 printf("uma_zalloc_arg: Bucketzone returned NULL\n"); 2282#endif 2283 2284zalloc_item: 2285 item = zone_alloc_item(zone, udata, flags); 2286 2287 return (item); 2288} 2289 2290static uma_slab_t 2291keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) 2292{ 2293 uma_slab_t slab; 2294 int reserve; 2295 2296 mtx_assert(&keg->uk_lock, MA_OWNED); 2297 slab = NULL; 2298 reserve = 0; 2299 if ((flags & M_USE_RESERVE) == 0) 2300 reserve = keg->uk_reserve; 2301 2302 for (;;) { 2303 /* 2304 * Find a slab with some space. Prefer slabs that are partially 2305 * used over those that are totally full. This helps to reduce 2306 * fragmentation. 2307 */ 2308 if (keg->uk_free > reserve) { 2309 if (!LIST_EMPTY(&keg->uk_part_slab)) { 2310 slab = LIST_FIRST(&keg->uk_part_slab); 2311 } else { 2312 slab = LIST_FIRST(&keg->uk_free_slab); 2313 LIST_REMOVE(slab, us_link); 2314 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, 2315 us_link); 2316 } 2317 MPASS(slab->us_keg == keg); 2318 return (slab); 2319 } 2320 2321 /* 2322 * M_NOVM means don't ask at all! 2323 */ 2324 if (flags & M_NOVM) 2325 break; 2326 2327 if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { 2328 keg->uk_flags |= UMA_ZFLAG_FULL; 2329 /* 2330 * If this is not a multi-zone, set the FULL bit. 2331 * Otherwise slab_multi() takes care of it. 2332 */ 2333 if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { 2334 zone->uz_flags |= UMA_ZFLAG_FULL; 2335 zone_log_warning(zone); 2336 } 2337 if (flags & M_NOWAIT) 2338 break; 2339 zone->uz_sleeps++; 2340 msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); 2341 continue; 2342 } 2343 slab = keg_alloc_slab(keg, zone, flags); 2344 /* 2345 * If we got a slab here it's safe to mark it partially used 2346 * and return. We assume that the caller is going to remove 2347 * at least one item. 2348 */ 2349 if (slab) { 2350 MPASS(slab->us_keg == keg); 2351 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2352 return (slab); 2353 } 2354 /* 2355 * We might not have been able to get a slab but another cpu 2356 * could have while we were unlocked. Check again before we 2357 * fail. 2358 */ 2359 flags |= M_NOVM; 2360 } 2361 return (slab); 2362} 2363 2364static uma_slab_t 2365zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) 2366{ 2367 uma_slab_t slab; 2368 2369 if (keg == NULL) { 2370 keg = zone_first_keg(zone); 2371 KEG_LOCK(keg); 2372 } 2373 2374 for (;;) { 2375 slab = keg_fetch_slab(keg, zone, flags); 2376 if (slab) 2377 return (slab); 2378 if (flags & (M_NOWAIT | M_NOVM)) 2379 break; 2380 } 2381 KEG_UNLOCK(keg); 2382 return (NULL); 2383} 2384 2385/* 2386 * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns 2387 * with the keg locked. On NULL no lock is held. 2388 * 2389 * The last pointer is used to seed the search. It is not required. 2390 */ 2391static uma_slab_t 2392zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) 2393{ 2394 uma_klink_t klink; 2395 uma_slab_t slab; 2396 uma_keg_t keg; 2397 int flags; 2398 int empty; 2399 int full; 2400 2401 /* 2402 * Don't wait on the first pass. This will skip limit tests 2403 * as well. We don't want to block if we can find a provider 2404 * without blocking. 2405 */ 2406 flags = (rflags & ~M_WAITOK) | M_NOWAIT; 2407 /* 2408 * Use the last slab allocated as a hint for where to start 2409 * the search. 2410 */ 2411 if (last != NULL) { 2412 slab = keg_fetch_slab(last, zone, flags); 2413 if (slab) 2414 return (slab); 2415 KEG_UNLOCK(last); 2416 } 2417 /* 2418 * Loop until we have a slab incase of transient failures 2419 * while M_WAITOK is specified. I'm not sure this is 100% 2420 * required but we've done it for so long now. 2421 */ 2422 for (;;) { 2423 empty = 0; 2424 full = 0; 2425 /* 2426 * Search the available kegs for slabs. Be careful to hold the 2427 * correct lock while calling into the keg layer. 2428 */ 2429 LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { 2430 keg = klink->kl_keg; 2431 KEG_LOCK(keg); 2432 if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { 2433 slab = keg_fetch_slab(keg, zone, flags); 2434 if (slab) 2435 return (slab); 2436 } 2437 if (keg->uk_flags & UMA_ZFLAG_FULL) 2438 full++; 2439 else 2440 empty++; 2441 KEG_UNLOCK(keg); 2442 } 2443 if (rflags & (M_NOWAIT | M_NOVM)) 2444 break; 2445 flags = rflags; 2446 /* 2447 * All kegs are full. XXX We can't atomically check all kegs 2448 * and sleep so just sleep for a short period and retry. 2449 */ 2450 if (full && !empty) { 2451 ZONE_LOCK(zone); 2452 zone->uz_flags |= UMA_ZFLAG_FULL; 2453 zone->uz_sleeps++; 2454 zone_log_warning(zone); 2455 msleep(zone, zone->uz_lockptr, PVM, 2456 "zonelimit", hz/100); 2457 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2458 ZONE_UNLOCK(zone); 2459 continue; 2460 } 2461 } 2462 return (NULL); 2463} 2464 2465static void * 2466slab_alloc_item(uma_keg_t keg, uma_slab_t slab) 2467{ 2468 void *item; 2469 uint8_t freei; 2470 2471 MPASS(keg == slab->us_keg); 2472 mtx_assert(&keg->uk_lock, MA_OWNED); 2473 2474 freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; 2475 BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); 2476 item = slab->us_data + (keg->uk_rsize * freei); 2477 slab->us_freecount--; 2478 keg->uk_free--; 2479 2480 /* Move this slab to the full list */ 2481 if (slab->us_freecount == 0) { 2482 LIST_REMOVE(slab, us_link); 2483 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); 2484 } 2485 2486 return (item); 2487} 2488 2489static int 2490zone_import(uma_zone_t zone, void **bucket, int max, int flags) 2491{ 2492 uma_slab_t slab; 2493 uma_keg_t keg; 2494 int i; 2495 2496 slab = NULL; 2497 keg = NULL; 2498 /* Try to keep the buckets totally full */ 2499 for (i = 0; i < max; ) { 2500 if ((slab = zone->uz_slab(zone, keg, flags)) == NULL) 2501 break; 2502 keg = slab->us_keg; 2503 while (slab->us_freecount && i < max) { 2504 bucket[i++] = slab_alloc_item(keg, slab); 2505 if (keg->uk_free <= keg->uk_reserve) 2506 break; 2507 } 2508 /* Don't grab more than one slab at a time. */ 2509 flags &= ~M_WAITOK; 2510 flags |= M_NOWAIT; 2511 } 2512 if (slab != NULL) 2513 KEG_UNLOCK(keg); 2514 2515 return i; 2516} 2517 2518static uma_bucket_t 2519zone_alloc_bucket(uma_zone_t zone, void *udata, int flags) 2520{ 2521 uma_bucket_t bucket; 2522 int max; 2523 2524 /* Don't wait for buckets, preserve caller's NOVM setting. */ 2525 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); 2526 if (bucket == NULL) 2527 return (NULL); 2528 2529 max = MIN(bucket->ub_entries, zone->uz_count); 2530 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, 2531 max, flags); 2532 2533 /* 2534 * Initialize the memory if necessary. 2535 */ 2536 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { 2537 int i; 2538 2539 for (i = 0; i < bucket->ub_cnt; i++) 2540 if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, 2541 flags) != 0) 2542 break; 2543 /* 2544 * If we couldn't initialize the whole bucket, put the 2545 * rest back onto the freelist. 2546 */ 2547 if (i != bucket->ub_cnt) { 2548 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], 2549 bucket->ub_cnt - i); 2550#ifdef INVARIANTS 2551 bzero(&bucket->ub_bucket[i], 2552 sizeof(void *) * (bucket->ub_cnt - i)); 2553#endif 2554 bucket->ub_cnt = i; 2555 } 2556 } 2557 2558 if (bucket->ub_cnt == 0) { 2559 bucket_free(zone, bucket, udata); 2560 atomic_add_long(&zone->uz_fails, 1); 2561 return (NULL); 2562 } 2563 2564 return (bucket); 2565} 2566 2567/* 2568 * Allocates a single item from a zone. 2569 * 2570 * Arguments 2571 * zone The zone to alloc for. 2572 * udata The data to be passed to the constructor. 2573 * flags M_WAITOK, M_NOWAIT, M_ZERO. 2574 * 2575 * Returns 2576 * NULL if there is no memory and M_NOWAIT is set 2577 * An item if successful 2578 */ 2579 2580static void * 2581zone_alloc_item(uma_zone_t zone, void *udata, int flags) 2582{ 2583 void *item; 2584 2585 item = NULL; 2586 2587#ifdef UMA_DEBUG_ALLOC 2588 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); 2589#endif 2590 if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1) 2591 goto fail; 2592 atomic_add_long(&zone->uz_allocs, 1); 2593 2594 /* 2595 * We have to call both the zone's init (not the keg's init) 2596 * and the zone's ctor. This is because the item is going from 2597 * a keg slab directly to the user, and the user is expecting it 2598 * to be both zone-init'd as well as zone-ctor'd. 2599 */ 2600 if (zone->uz_init != NULL) { 2601 if (zone->uz_init(item, zone->uz_size, flags) != 0) { 2602 zone_free_item(zone, item, udata, SKIP_FINI); 2603 goto fail; 2604 } 2605 } 2606 if (zone->uz_ctor != NULL) { 2607 if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { 2608 zone_free_item(zone, item, udata, SKIP_DTOR); 2609 goto fail; 2610 } 2611 } 2612#ifdef INVARIANTS 2613 uma_dbg_alloc(zone, NULL, item); 2614#endif 2615 if (flags & M_ZERO) 2616 bzero(item, zone->uz_size); 2617 2618 return (item); 2619 2620fail: 2621 atomic_add_long(&zone->uz_fails, 1); 2622 return (NULL); 2623} 2624 2625/* See uma.h */ 2626void 2627uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 2628{ 2629 uma_cache_t cache; 2630 uma_bucket_t bucket; 2631 int lockfail; 2632 int cpu; 2633 2634#ifdef UMA_DEBUG_ALLOC_1 2635 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); 2636#endif 2637 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, 2638 zone->uz_name); 2639 2640 /* uma_zfree(..., NULL) does nothing, to match free(9). */ 2641 if (item == NULL) 2642 return; 2643#ifdef DEBUG_MEMGUARD 2644 if (is_memguard_addr(item)) { 2645 if (zone->uz_dtor != NULL && zone->uz_dtor != mtrash_dtor) 2646 zone->uz_dtor(item, zone->uz_size, udata); 2647 if (zone->uz_fini != NULL && zone->uz_fini != mtrash_fini) 2648 zone->uz_fini(item, zone->uz_size); 2649 memguard_free(item); 2650 return; 2651 } 2652#endif 2653#ifdef INVARIANTS 2654 if (zone->uz_flags & UMA_ZONE_MALLOC) 2655 uma_dbg_free(zone, udata, item); 2656 else 2657 uma_dbg_free(zone, NULL, item); 2658#endif 2659 if (zone->uz_dtor != NULL) 2660 zone->uz_dtor(item, zone->uz_size, udata); 2661 2662 /* 2663 * The race here is acceptable. If we miss it we'll just have to wait 2664 * a little longer for the limits to be reset. 2665 */ 2666 if (zone->uz_flags & UMA_ZFLAG_FULL) 2667 goto zfree_item; 2668 2669 /* 2670 * If possible, free to the per-CPU cache. There are two 2671 * requirements for safe access to the per-CPU cache: (1) the thread 2672 * accessing the cache must not be preempted or yield during access, 2673 * and (2) the thread must not migrate CPUs without switching which 2674 * cache it accesses. We rely on a critical section to prevent 2675 * preemption and migration. We release the critical section in 2676 * order to acquire the zone mutex if we are unable to free to the 2677 * current cache; when we re-acquire the critical section, we must 2678 * detect and handle migration if it has occurred. 2679 */ 2680zfree_restart: 2681 critical_enter(); 2682 cpu = curcpu; 2683 cache = &zone->uz_cpu[cpu]; 2684 2685zfree_start: 2686 /* 2687 * Try to free into the allocbucket first to give LIFO ordering 2688 * for cache-hot datastructures. Spill over into the freebucket 2689 * if necessary. Alloc will swap them if one runs dry. 2690 */ 2691 bucket = cache->uc_allocbucket; 2692 if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) 2693 bucket = cache->uc_freebucket; 2694 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2695 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 2696 ("uma_zfree: Freeing to non free bucket index.")); 2697 bucket->ub_bucket[bucket->ub_cnt] = item; 2698 bucket->ub_cnt++; 2699 cache->uc_frees++; 2700 critical_exit(); 2701 return; 2702 } 2703 2704 /* 2705 * We must go back the zone, which requires acquiring the zone lock, 2706 * which in turn means we must release and re-acquire the critical 2707 * section. Since the critical section is released, we may be 2708 * preempted or migrate. As such, make sure not to maintain any 2709 * thread-local state specific to the cache from prior to releasing 2710 * the critical section. 2711 */ 2712 critical_exit(); 2713 if (zone->uz_count == 0 || bucketdisable) 2714 goto zfree_item; 2715 2716 lockfail = 0; 2717 if (ZONE_TRYLOCK(zone) == 0) { 2718 /* Record contention to size the buckets. */ 2719 ZONE_LOCK(zone); 2720 lockfail = 1; 2721 } 2722 critical_enter(); 2723 cpu = curcpu; 2724 cache = &zone->uz_cpu[cpu]; 2725 2726 /* 2727 * Since we have locked the zone we may as well send back our stats. 2728 */ 2729 atomic_add_long(&zone->uz_allocs, cache->uc_allocs); 2730 atomic_add_long(&zone->uz_frees, cache->uc_frees); 2731 cache->uc_allocs = 0; 2732 cache->uc_frees = 0; 2733 2734 bucket = cache->uc_freebucket; 2735 if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { 2736 ZONE_UNLOCK(zone); 2737 goto zfree_start; 2738 } 2739 cache->uc_freebucket = NULL; 2740 2741 /* Can we throw this on the zone full list? */ 2742 if (bucket != NULL) { 2743#ifdef UMA_DEBUG_ALLOC 2744 printf("uma_zfree: Putting old bucket on the free list.\n"); 2745#endif 2746 /* ub_cnt is pointing to the last free item */ 2747 KASSERT(bucket->ub_cnt != 0, 2748 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 2749 LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); 2750 } 2751 2752 /* We are no longer associated with this CPU. */ 2753 critical_exit(); 2754 2755 /* 2756 * We bump the uz count when the cache size is insufficient to 2757 * handle the working set. 2758 */ 2759 if (lockfail && zone->uz_count < BUCKET_MAX) 2760 zone->uz_count++; 2761 ZONE_UNLOCK(zone); 2762 2763#ifdef UMA_DEBUG_ALLOC 2764 printf("uma_zfree: Allocating new free bucket.\n"); 2765#endif 2766 bucket = bucket_alloc(zone, udata, M_NOWAIT); 2767 if (bucket) { 2768 critical_enter(); 2769 cpu = curcpu; 2770 cache = &zone->uz_cpu[cpu]; 2771 if (cache->uc_freebucket == NULL) { 2772 cache->uc_freebucket = bucket; 2773 goto zfree_start; 2774 } 2775 /* 2776 * We lost the race, start over. We have to drop our 2777 * critical section to free the bucket. 2778 */ 2779 critical_exit(); 2780 bucket_free(zone, bucket, udata); 2781 goto zfree_restart; 2782 } 2783 2784 /* 2785 * If nothing else caught this, we'll just do an internal free. 2786 */ 2787zfree_item: 2788 zone_free_item(zone, item, udata, SKIP_DTOR); 2789 2790 return; 2791} 2792 2793static void 2794slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) 2795{ 2796 uint8_t freei; 2797 2798 mtx_assert(&keg->uk_lock, MA_OWNED); 2799 MPASS(keg == slab->us_keg); 2800 2801 /* Do we need to remove from any lists? */ 2802 if (slab->us_freecount+1 == keg->uk_ipers) { 2803 LIST_REMOVE(slab, us_link); 2804 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 2805 } else if (slab->us_freecount == 0) { 2806 LIST_REMOVE(slab, us_link); 2807 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); 2808 } 2809 2810 /* Slab management. */ 2811 freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 2812 BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); 2813 slab->us_freecount++; 2814 2815 /* Keg statistics. */ 2816 keg->uk_free++; 2817} 2818 2819static void 2820zone_release(uma_zone_t zone, void **bucket, int cnt) 2821{ 2822 void *item; 2823 uma_slab_t slab; 2824 uma_keg_t keg; 2825 uint8_t *mem; 2826 int clearfull; 2827 int i; 2828 2829 clearfull = 0; 2830 keg = zone_first_keg(zone); 2831 KEG_LOCK(keg); 2832 for (i = 0; i < cnt; i++) { 2833 item = bucket[i]; 2834 if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { 2835 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); 2836 if (zone->uz_flags & UMA_ZONE_HASH) { 2837 slab = hash_sfind(&keg->uk_hash, mem); 2838 } else { 2839 mem += keg->uk_pgoff; 2840 slab = (uma_slab_t)mem; 2841 } 2842 } else { 2843 slab = vtoslab((vm_offset_t)item); 2844 if (slab->us_keg != keg) { 2845 KEG_UNLOCK(keg); 2846 keg = slab->us_keg; 2847 KEG_LOCK(keg); 2848 } 2849 } 2850 slab_free_item(keg, slab, item); 2851 if (keg->uk_flags & UMA_ZFLAG_FULL) { 2852 if (keg->uk_pages < keg->uk_maxpages) { 2853 keg->uk_flags &= ~UMA_ZFLAG_FULL; 2854 clearfull = 1; 2855 } 2856 2857 /* 2858 * We can handle one more allocation. Since we're 2859 * clearing ZFLAG_FULL, wake up all procs blocked 2860 * on pages. This should be uncommon, so keeping this 2861 * simple for now (rather than adding count of blocked 2862 * threads etc). 2863 */ 2864 wakeup(keg); 2865 } 2866 } 2867 KEG_UNLOCK(keg); 2868 if (clearfull) { 2869 ZONE_LOCK(zone); 2870 zone->uz_flags &= ~UMA_ZFLAG_FULL; 2871 wakeup(zone); 2872 ZONE_UNLOCK(zone); 2873 } 2874 2875} 2876 2877/* 2878 * Frees a single item to any zone. 2879 * 2880 * Arguments: 2881 * zone The zone to free to 2882 * item The item we're freeing 2883 * udata User supplied data for the dtor 2884 * skip Skip dtors and finis 2885 */ 2886static void 2887zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) 2888{ 2889 2890#ifdef INVARIANTS 2891 if (skip == SKIP_NONE) { 2892 if (zone->uz_flags & UMA_ZONE_MALLOC) 2893 uma_dbg_free(zone, udata, item); 2894 else 2895 uma_dbg_free(zone, NULL, item); 2896 } 2897#endif 2898 if (skip < SKIP_DTOR && zone->uz_dtor) 2899 zone->uz_dtor(item, zone->uz_size, udata); 2900 2901 if (skip < SKIP_FINI && zone->uz_fini) 2902 zone->uz_fini(item, zone->uz_size); 2903 2904 atomic_add_long(&zone->uz_frees, 1); 2905 zone->uz_release(zone->uz_arg, &item, 1); 2906} 2907 2908/* See uma.h */ 2909int 2910uma_zone_set_max(uma_zone_t zone, int nitems) 2911{ 2912 uma_keg_t keg; 2913 2914 keg = zone_first_keg(zone); 2915 if (keg == NULL) 2916 return (0); 2917 KEG_LOCK(keg); 2918 keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; 2919 if (keg->uk_maxpages * keg->uk_ipers < nitems) 2920 keg->uk_maxpages += keg->uk_ppera; 2921 nitems = keg->uk_maxpages * keg->uk_ipers; 2922 KEG_UNLOCK(keg); 2923 2924 return (nitems); 2925} 2926 2927/* See uma.h */ 2928int 2929uma_zone_get_max(uma_zone_t zone) 2930{ 2931 int nitems; 2932 uma_keg_t keg; 2933 2934 keg = zone_first_keg(zone); 2935 if (keg == NULL) 2936 return (0); 2937 KEG_LOCK(keg); 2938 nitems = keg->uk_maxpages * keg->uk_ipers; 2939 KEG_UNLOCK(keg); 2940 2941 return (nitems); 2942} 2943 2944/* See uma.h */ 2945void 2946uma_zone_set_warning(uma_zone_t zone, const char *warning) 2947{ 2948 2949 ZONE_LOCK(zone); 2950 zone->uz_warning = warning; 2951 ZONE_UNLOCK(zone); 2952} 2953 2954/* See uma.h */ 2955int 2956uma_zone_get_cur(uma_zone_t zone) 2957{ 2958 int64_t nitems; 2959 u_int i; 2960 2961 ZONE_LOCK(zone); 2962 nitems = zone->uz_allocs - zone->uz_frees; 2963 CPU_FOREACH(i) { 2964 /* 2965 * See the comment in sysctl_vm_zone_stats() regarding the 2966 * safety of accessing the per-cpu caches. With the zone lock 2967 * held, it is safe, but can potentially result in stale data. 2968 */ 2969 nitems += zone->uz_cpu[i].uc_allocs - 2970 zone->uz_cpu[i].uc_frees; 2971 } 2972 ZONE_UNLOCK(zone); 2973 2974 return (nitems < 0 ? 0 : nitems); 2975} 2976 2977/* See uma.h */ 2978void 2979uma_zone_set_init(uma_zone_t zone, uma_init uminit) 2980{ 2981 uma_keg_t keg; 2982 2983 keg = zone_first_keg(zone); 2984 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 2985 KEG_LOCK(keg); 2986 KASSERT(keg->uk_pages == 0, 2987 ("uma_zone_set_init on non-empty keg")); 2988 keg->uk_init = uminit; 2989 KEG_UNLOCK(keg); 2990} 2991 2992/* See uma.h */ 2993void 2994uma_zone_set_fini(uma_zone_t zone, uma_fini fini) 2995{ 2996 uma_keg_t keg; 2997 2998 keg = zone_first_keg(zone); 2999 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 3000 KEG_LOCK(keg); 3001 KASSERT(keg->uk_pages == 0, 3002 ("uma_zone_set_fini on non-empty keg")); 3003 keg->uk_fini = fini; 3004 KEG_UNLOCK(keg); 3005} 3006 3007/* See uma.h */ 3008void 3009uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) 3010{ 3011 3012 ZONE_LOCK(zone); 3013 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3014 ("uma_zone_set_zinit on non-empty keg")); 3015 zone->uz_init = zinit; 3016 ZONE_UNLOCK(zone); 3017} 3018 3019/* See uma.h */ 3020void 3021uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) 3022{ 3023 3024 ZONE_LOCK(zone); 3025 KASSERT(zone_first_keg(zone)->uk_pages == 0, 3026 ("uma_zone_set_zfini on non-empty keg")); 3027 zone->uz_fini = zfini; 3028 ZONE_UNLOCK(zone); 3029} 3030 3031/* See uma.h */ 3032/* XXX uk_freef is not actually used with the zone locked */ 3033void 3034uma_zone_set_freef(uma_zone_t zone, uma_free freef) 3035{ 3036 uma_keg_t keg; 3037 3038 keg = zone_first_keg(zone); 3039 KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); 3040 KEG_LOCK(keg); 3041 keg->uk_freef = freef; 3042 KEG_UNLOCK(keg); 3043} 3044 3045/* See uma.h */ 3046/* XXX uk_allocf is not actually used with the zone locked */ 3047void 3048uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 3049{ 3050 uma_keg_t keg; 3051 3052 keg = zone_first_keg(zone); 3053 KEG_LOCK(keg); 3054 keg->uk_allocf = allocf; 3055 KEG_UNLOCK(keg); 3056} 3057 3058/* See uma.h */ 3059void 3060uma_zone_reserve(uma_zone_t zone, int items) 3061{ 3062 uma_keg_t keg; 3063 3064 keg = zone_first_keg(zone); 3065 if (keg == NULL) 3066 return; 3067 KEG_LOCK(keg); 3068 keg->uk_reserve = items; 3069 KEG_UNLOCK(keg); 3070 3071 return; 3072} 3073 3074/* See uma.h */ 3075int 3076uma_zone_reserve_kva(uma_zone_t zone, int count) 3077{ 3078 uma_keg_t keg; 3079 vm_offset_t kva; 3080 int pages; 3081 3082 keg = zone_first_keg(zone); 3083 if (keg == NULL) 3084 return (0); 3085 pages = count / keg->uk_ipers; 3086 3087 if (pages * keg->uk_ipers < count) 3088 pages++; 3089 3090#ifdef UMA_MD_SMALL_ALLOC 3091 if (keg->uk_ppera > 1) { 3092#else 3093 if (1) { 3094#endif 3095 kva = kva_alloc(pages * UMA_SLAB_SIZE); 3096 if (kva == 0) 3097 return (0); 3098 } else 3099 kva = 0; 3100 KEG_LOCK(keg); 3101 keg->uk_kva = kva; 3102 keg->uk_offset = 0; 3103 keg->uk_maxpages = pages; 3104#ifdef UMA_MD_SMALL_ALLOC 3105 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; 3106#else 3107 keg->uk_allocf = noobj_alloc; 3108#endif 3109 keg->uk_flags |= UMA_ZONE_NOFREE; 3110 KEG_UNLOCK(keg); 3111 3112 return (1); 3113} 3114 3115/* See uma.h */ 3116void 3117uma_prealloc(uma_zone_t zone, int items) 3118{ 3119 int slabs; 3120 uma_slab_t slab; 3121 uma_keg_t keg; 3122 3123 keg = zone_first_keg(zone); 3124 if (keg == NULL) 3125 return; 3126 KEG_LOCK(keg); 3127 slabs = items / keg->uk_ipers; 3128 if (slabs * keg->uk_ipers < items) 3129 slabs++; 3130 while (slabs > 0) { 3131 slab = keg_alloc_slab(keg, zone, M_WAITOK); 3132 if (slab == NULL) 3133 break; 3134 MPASS(slab->us_keg == keg); 3135 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); 3136 slabs--; 3137 } 3138 KEG_UNLOCK(keg); 3139} 3140 3141/* See uma.h */ 3142uint32_t * 3143uma_find_refcnt(uma_zone_t zone, void *item) 3144{ 3145 uma_slabrefcnt_t slabref; 3146 uma_slab_t slab; 3147 uma_keg_t keg; 3148 uint32_t *refcnt; 3149 int idx; 3150 3151 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK)); 3152 slabref = (uma_slabrefcnt_t)slab; 3153 keg = slab->us_keg; 3154 KASSERT(keg->uk_flags & UMA_ZONE_REFCNT, 3155 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT")); 3156 idx = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; 3157 refcnt = &slabref->us_refcnt[idx]; 3158 return refcnt; 3159} 3160 3161/* See uma.h */ 3162void 3163uma_reclaim(void) 3164{ 3165#ifdef UMA_DEBUG 3166 printf("UMA: vm asked us to release pages!\n"); 3167#endif 3168 bucket_enable(); 3169 zone_foreach(zone_drain); 3170 if (vm_page_count_min()) { 3171 cache_drain_safe(NULL); 3172 zone_foreach(zone_drain); 3173 } 3174 /* 3175 * Some slabs may have been freed but this zone will be visited early 3176 * we visit again so that we can free pages that are empty once other 3177 * zones are drained. We have to do the same for buckets. 3178 */ 3179 zone_drain(slabzone); 3180 zone_drain(slabrefzone); 3181 bucket_zone_drain(); 3182} 3183 3184/* See uma.h */ 3185int 3186uma_zone_exhausted(uma_zone_t zone) 3187{ 3188 int full; 3189 3190 ZONE_LOCK(zone); 3191 full = (zone->uz_flags & UMA_ZFLAG_FULL); 3192 ZONE_UNLOCK(zone); 3193 return (full); 3194} 3195 3196int 3197uma_zone_exhausted_nolock(uma_zone_t zone) 3198{ 3199 return (zone->uz_flags & UMA_ZFLAG_FULL); 3200} 3201 3202void * 3203uma_large_malloc(int size, int wait) 3204{ 3205 void *mem; 3206 uma_slab_t slab; 3207 uint8_t flags; 3208 3209 slab = zone_alloc_item(slabzone, NULL, wait); 3210 if (slab == NULL) 3211 return (NULL); 3212 mem = page_alloc(NULL, size, &flags, wait); 3213 if (mem) { 3214 vsetslab((vm_offset_t)mem, slab); 3215 slab->us_data = mem; 3216 slab->us_flags = flags | UMA_SLAB_MALLOC; 3217 slab->us_size = size; 3218 } else { 3219 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3220 } 3221 3222 return (mem); 3223} 3224 3225void 3226uma_large_free(uma_slab_t slab) 3227{ 3228 3229 page_free(slab->us_data, slab->us_size, slab->us_flags); 3230 zone_free_item(slabzone, slab, NULL, SKIP_NONE); 3231} 3232 3233void 3234uma_print_stats(void) 3235{ 3236 zone_foreach(uma_print_zone); 3237} 3238 3239static void 3240slab_print(uma_slab_t slab) 3241{ 3242 printf("slab: keg %p, data %p, freecount %d\n", 3243 slab->us_keg, slab->us_data, slab->us_freecount); 3244} 3245 3246static void 3247cache_print(uma_cache_t cache) 3248{ 3249 printf("alloc: %p(%d), free: %p(%d)\n", 3250 cache->uc_allocbucket, 3251 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, 3252 cache->uc_freebucket, 3253 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); 3254} 3255 3256static void 3257uma_print_keg(uma_keg_t keg) 3258{ 3259 uma_slab_t slab; 3260 3261 printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " 3262 "out %d free %d limit %d\n", 3263 keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, 3264 keg->uk_ipers, keg->uk_ppera, 3265 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free, 3266 (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); 3267 printf("Part slabs:\n"); 3268 LIST_FOREACH(slab, &keg->uk_part_slab, us_link) 3269 slab_print(slab); 3270 printf("Free slabs:\n"); 3271 LIST_FOREACH(slab, &keg->uk_free_slab, us_link) 3272 slab_print(slab); 3273 printf("Full slabs:\n"); 3274 LIST_FOREACH(slab, &keg->uk_full_slab, us_link) 3275 slab_print(slab); 3276} 3277 3278void 3279uma_print_zone(uma_zone_t zone) 3280{ 3281 uma_cache_t cache; 3282 uma_klink_t kl; 3283 int i; 3284 3285 printf("zone: %s(%p) size %d flags %#x\n", 3286 zone->uz_name, zone, zone->uz_size, zone->uz_flags); 3287 LIST_FOREACH(kl, &zone->uz_kegs, kl_link) 3288 uma_print_keg(kl->kl_keg); 3289 CPU_FOREACH(i) { 3290 cache = &zone->uz_cpu[i]; 3291 printf("CPU %d Cache:\n", i); 3292 cache_print(cache); 3293 } 3294} 3295 3296#ifdef DDB 3297/* 3298 * Generate statistics across both the zone and its per-cpu cache's. Return 3299 * desired statistics if the pointer is non-NULL for that statistic. 3300 * 3301 * Note: does not update the zone statistics, as it can't safely clear the 3302 * per-CPU cache statistic. 3303 * 3304 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't 3305 * safe from off-CPU; we should modify the caches to track this information 3306 * directly so that we don't have to. 3307 */ 3308static void 3309uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, 3310 uint64_t *freesp, uint64_t *sleepsp) 3311{ 3312 uma_cache_t cache; 3313 uint64_t allocs, frees, sleeps; 3314 int cachefree, cpu; 3315 3316 allocs = frees = sleeps = 0; 3317 cachefree = 0; 3318 CPU_FOREACH(cpu) { 3319 cache = &z->uz_cpu[cpu]; 3320 if (cache->uc_allocbucket != NULL) 3321 cachefree += cache->uc_allocbucket->ub_cnt; 3322 if (cache->uc_freebucket != NULL) 3323 cachefree += cache->uc_freebucket->ub_cnt; 3324 allocs += cache->uc_allocs; 3325 frees += cache->uc_frees; 3326 } 3327 allocs += z->uz_allocs; 3328 frees += z->uz_frees; 3329 sleeps += z->uz_sleeps; 3330 if (cachefreep != NULL) 3331 *cachefreep = cachefree; 3332 if (allocsp != NULL) 3333 *allocsp = allocs; 3334 if (freesp != NULL) 3335 *freesp = frees; 3336 if (sleepsp != NULL) 3337 *sleepsp = sleeps; 3338} 3339#endif /* DDB */ 3340 3341static int 3342sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) 3343{ 3344 uma_keg_t kz; 3345 uma_zone_t z; 3346 int count; 3347 3348 count = 0; 3349 mtx_lock(&uma_mtx); 3350 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3351 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3352 count++; 3353 } 3354 mtx_unlock(&uma_mtx); 3355 return (sysctl_handle_int(oidp, &count, 0, req)); 3356} 3357 3358static int 3359sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) 3360{ 3361 struct uma_stream_header ush; 3362 struct uma_type_header uth; 3363 struct uma_percpu_stat ups; 3364 uma_bucket_t bucket; 3365 struct sbuf sbuf; 3366 uma_cache_t cache; 3367 uma_klink_t kl; 3368 uma_keg_t kz; 3369 uma_zone_t z; 3370 uma_keg_t k; 3371 int count, error, i; 3372 3373 error = sysctl_wire_old_buffer(req, 0); 3374 if (error != 0) 3375 return (error); 3376 sbuf_new_for_sysctl(&sbuf, NULL, 128, req); 3377 3378 count = 0; 3379 mtx_lock(&uma_mtx); 3380 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3381 LIST_FOREACH(z, &kz->uk_zones, uz_link) 3382 count++; 3383 } 3384 3385 /* 3386 * Insert stream header. 3387 */ 3388 bzero(&ush, sizeof(ush)); 3389 ush.ush_version = UMA_STREAM_VERSION; 3390 ush.ush_maxcpus = (mp_maxid + 1); 3391 ush.ush_count = count; 3392 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); 3393 3394 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3395 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3396 bzero(&uth, sizeof(uth)); 3397 ZONE_LOCK(z); 3398 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); 3399 uth.uth_align = kz->uk_align; 3400 uth.uth_size = kz->uk_size; 3401 uth.uth_rsize = kz->uk_rsize; 3402 LIST_FOREACH(kl, &z->uz_kegs, kl_link) { 3403 k = kl->kl_keg; 3404 uth.uth_maxpages += k->uk_maxpages; 3405 uth.uth_pages += k->uk_pages; 3406 uth.uth_keg_free += k->uk_free; 3407 uth.uth_limit = (k->uk_maxpages / k->uk_ppera) 3408 * k->uk_ipers; 3409 } 3410 3411 /* 3412 * A zone is secondary is it is not the first entry 3413 * on the keg's zone list. 3414 */ 3415 if ((z->uz_flags & UMA_ZONE_SECONDARY) && 3416 (LIST_FIRST(&kz->uk_zones) != z)) 3417 uth.uth_zone_flags = UTH_ZONE_SECONDARY; 3418 3419 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3420 uth.uth_zone_free += bucket->ub_cnt; 3421 uth.uth_allocs = z->uz_allocs; 3422 uth.uth_frees = z->uz_frees; 3423 uth.uth_fails = z->uz_fails; 3424 uth.uth_sleeps = z->uz_sleeps; 3425 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); 3426 /* 3427 * While it is not normally safe to access the cache 3428 * bucket pointers while not on the CPU that owns the 3429 * cache, we only allow the pointers to be exchanged 3430 * without the zone lock held, not invalidated, so 3431 * accept the possible race associated with bucket 3432 * exchange during monitoring. 3433 */ 3434 for (i = 0; i < (mp_maxid + 1); i++) { 3435 bzero(&ups, sizeof(ups)); 3436 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) 3437 goto skip; 3438 if (CPU_ABSENT(i)) 3439 goto skip; 3440 cache = &z->uz_cpu[i]; 3441 if (cache->uc_allocbucket != NULL) 3442 ups.ups_cache_free += 3443 cache->uc_allocbucket->ub_cnt; 3444 if (cache->uc_freebucket != NULL) 3445 ups.ups_cache_free += 3446 cache->uc_freebucket->ub_cnt; 3447 ups.ups_allocs = cache->uc_allocs; 3448 ups.ups_frees = cache->uc_frees; 3449skip: 3450 (void)sbuf_bcat(&sbuf, &ups, sizeof(ups)); 3451 } 3452 ZONE_UNLOCK(z); 3453 } 3454 } 3455 mtx_unlock(&uma_mtx); 3456 error = sbuf_finish(&sbuf); 3457 sbuf_delete(&sbuf); 3458 return (error); 3459} 3460 3461#ifdef DDB 3462DB_SHOW_COMMAND(uma, db_show_uma) 3463{ 3464 uint64_t allocs, frees, sleeps; 3465 uma_bucket_t bucket; 3466 uma_keg_t kz; 3467 uma_zone_t z; 3468 int cachefree; 3469 3470 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", 3471 "Requests", "Sleeps"); 3472 LIST_FOREACH(kz, &uma_kegs, uk_link) { 3473 LIST_FOREACH(z, &kz->uk_zones, uz_link) { 3474 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { 3475 allocs = z->uz_allocs; 3476 frees = z->uz_frees; 3477 sleeps = z->uz_sleeps; 3478 cachefree = 0; 3479 } else 3480 uma_zone_sumstat(z, &cachefree, &allocs, 3481 &frees, &sleeps); 3482 if (!((z->uz_flags & UMA_ZONE_SECONDARY) && 3483 (LIST_FIRST(&kz->uk_zones) != z))) 3484 cachefree += kz->uk_free; 3485 LIST_FOREACH(bucket, &z->uz_buckets, ub_link) 3486 cachefree += bucket->ub_cnt; 3487 db_printf("%18s %8ju %8jd %8d %12ju %8ju\n", z->uz_name, 3488 (uintmax_t)kz->uk_size, 3489 (intmax_t)(allocs - frees), cachefree, 3490 (uintmax_t)allocs, sleeps); 3491 if (db_pager_quit) 3492 return; 3493 } 3494 } 3495} 3496#endif 3497